Initial commit

fbshipit-source-id: da6be2f26e3a1202f4bffde8cb980e2dcb851294

sam3/model/__init__.py

@@ -0,0 +1 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved

sam3/model/act_ckpt_utils.py

@@ -0,0 +1,114 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import inspect
+from functools import wraps
+from typing import Callable, TypeVar, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint as checkpoint
+from torch.utils._pytree import tree_map_only
+
+# Type variables for better type hinting
+T = TypeVar("T")
+Module = TypeVar("Module", bound=nn.Module)
+
+
+def activation_ckpt_wrapper(module: Union[nn.Module, Callable]) -> Callable:
+    """
+    Wraps a given module to enable or disable activation checkpointing.
+
+    Activation checkpointing (gradient checkpointing) trades compute for memory by
+    recomputing intermediate activations during the backward pass instead of storing
+    them in memory during the forward pass.
+
+    When activation checkpointing is enabled, the wrapper expects only keyword arguments,
+    and it maps these to positional arguments based on the module's signature.
+
+    Args:
+        module: The module or function to wrap with activation checkpointing
+
+    Returns:
+        A wrapped callable that supports activation checkpointing
+
+    Usage:
+        The returned wrapper function can be called with the same arguments as the
+        original module, with an additional `act_ckpt_enable` keyword argument to control
+        activation checkpointing and optional `use_reentrant` parameter.
+
+    Example:
+        ```python
+        wrapped_module = activation_ckpt_wrapper(my_module)
+        output = wrapped_module(x=input_tensor, y=another_tensor, act_ckpt_enable=True)
+        ```
+    """
+
+    @wraps(module)
+    def act_ckpt_wrapper(
+        *args, act_ckpt_enable: bool = True, use_reentrant: bool = False, **kwargs
+    ):
+        if act_ckpt_enable:
+            if len(args) > 0:
+                raise ValueError(
+                    "This wrapper expects keyword arguments only when `act_ckpt_enable=True`"
+                )
+            # Get the signature of the target function/module
+            callable_fn = module.forward if isinstance(module, nn.Module) else module
+            sig = inspect.signature(callable_fn)
+            # Create a mapping of parameter names to their default values
+            param_defaults = {
+                name: param.default for name, param in sig.parameters.items()
+            }
+            args = []
+            for p_name in param_defaults.keys():
+                if p_name in kwargs:
+                    args.append(kwargs.pop(p_name))
+                elif param_defaults[p_name] is not inspect.Parameter.empty:
+                    # Set arg to default value if it's not in kwargs. Useful for primitive types or args that default to None
+                    args.append(param_defaults[p_name])
+                elif (
+                    sig.parameters[p_name].kind is not inspect.Parameter.VAR_KEYWORD
+                ):  # Skip **kwargs parameter
+                    raise ValueError(f"Missing positional argument: {p_name}")
+
+            # Scan remaining kwargs for torch.Tensor
+            remaining_keys = list(kwargs.keys())
+            for key in remaining_keys:
+                if isinstance(kwargs[key], torch.Tensor):
+                    # Remove the tensor from kwargs, assuming it's not required by the module.
+                    # If it is required, the module's signature should be modified to accept it as a positional or keyword argument.
+                    kwargs[key] = "_REMOVED_BY_ACT_CKPT_WRAPPER_"
+
+            ret = checkpoint.checkpoint(
+                module, *args, use_reentrant=use_reentrant, **kwargs
+            )
+        else:
+            ret = module(*args, **kwargs)
+
+        return ret
+
+    return act_ckpt_wrapper
+
+
+def clone_output_wrapper(f: Callable[..., T]) -> Callable[..., T]:
+    """
+    Clone the CUDA output tensors of a function to avoid in-place operations.
+
+    This wrapper is useful when working with torch.compile to prevent errors
+    related to in-place operations on tensors.
+
+    Args:
+        f: The function whose CUDA tensor outputs should be cloned
+
+    Returns:
+        A wrapped function that clones any CUDA tensor outputs
+    """
+
+    @wraps(f)
+    def wrapped(*args, **kwargs):
+        outputs = f(*args, **kwargs)
+        return tree_map_only(
+            torch.Tensor, lambda t: t.clone() if t.is_cuda else t, outputs
+        )
+
+    return wrapped

sam3/model/box_ops.py

@@ -0,0 +1,217 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+"""
+Utilities for bounding box manipulation and GIoU.
+"""
+
+from typing import Tuple
+
+import torch
+
+
+def box_cxcywh_to_xyxy(x):
+    x_c, y_c, w, h = x.unbind(-1)
+    b = [(x_c - 0.5 * w), (y_c - 0.5 * h), (x_c + 0.5 * w), (y_c + 0.5 * h)]
+    return torch.stack(b, dim=-1)
+
+
+def box_cxcywh_to_xywh(x):
+    x_c, y_c, w, h = x.unbind(-1)
+    b = [(x_c - 0.5 * w), (y_c - 0.5 * h), (w), (h)]
+    return torch.stack(b, dim=-1)
+
+
+def box_xywh_to_xyxy(x):
+    x, y, w, h = x.unbind(-1)
+    b = [(x), (y), (x + w), (y + h)]
+    return torch.stack(b, dim=-1)
+
+
+def box_xywh_to_cxcywh(x):
+    x, y, w, h = x.unbind(-1)
+    b = [(x + 0.5 * w), (y + 0.5 * h), (w), (h)]
+    return torch.stack(b, dim=-1)
+
+
+def box_xyxy_to_xywh(x):
+    x, y, X, Y = x.unbind(-1)
+    b = [(x), (y), (X - x), (Y - y)]
+    return torch.stack(b, dim=-1)
+
+
+def box_xyxy_to_cxcywh(x):
+    x0, y0, x1, y1 = x.unbind(-1)
+    b = [(x0 + x1) / 2, (y0 + y1) / 2, (x1 - x0), (y1 - y0)]
+    return torch.stack(b, dim=-1)
+
+
+def box_area(boxes):
+    """
+    Batched version of box area. Boxes should be in [x0, y0, x1, y1] format.
+
+    Inputs:
+    - boxes: Tensor of shape (..., 4)
+
+    Returns:
+    - areas: Tensor of shape (...,)
+    """
+    x0, y0, x1, y1 = boxes.unbind(-1)
+    return (x1 - x0) * (y1 - y0)
+
+
+def masks_to_boxes(masks):
+    """Compute the bounding boxes around the provided masks
+
+    The masks should be in format [N, H, W] where N is the number of masks, (H, W) are the spatial dimensions.
+
+    Returns a [N, 4] tensors, with the boxes in xyxy format
+    """
+    if masks.numel() == 0:
+        return torch.zeros((0, 4), device=masks.device)
+
+    h, w = masks.shape[-2:]
+
+    y = torch.arange(0, h, dtype=torch.float, device=masks.device)
+    x = torch.arange(0, w, dtype=torch.float, device=masks.device)
+    y, x = torch.meshgrid(y, x)
+
+    x_mask = masks * x.unsqueeze(0)
+    x_max = x_mask.flatten(1).max(-1)[0] + 1
+    x_min = x_mask.masked_fill(~(masks.bool()), 1e8).flatten(1).min(-1)[0]
+
+    y_mask = masks * y.unsqueeze(0)
+    y_max = y_mask.flatten(1).max(-1)[0] + 1
+    y_min = y_mask.masked_fill(~(masks.bool()), 1e8).flatten(1).min(-1)[0]
+
+    boxes = torch.stack([x_min, y_min, x_max, y_max], 1)
+    # Invalidate boxes corresponding to empty masks.
+    boxes = boxes * masks.flatten(-2).any(-1)
+    return boxes
+
+
+def box_iou(boxes1, boxes2):
+    """
+    Batched version of box_iou. Boxes should be in [x0, y0, x1, y1] format.
+
+    Inputs:
+    - boxes1: Tensor of shape (..., N, 4)
+    - boxes2: Tensor of shape (..., M, 4)
+
+    Returns:
+    - iou, union: Tensors of shape (..., N, M)
+    """
+    area1 = box_area(boxes1)
+    area2 = box_area(boxes2)
+
+    # boxes1: (..., N, 4) -> (..., N, 1, 2)
+    # boxes2: (..., M, 4) -> (..., 1, M, 2)
+    lt = torch.max(boxes1[..., :, None, :2], boxes2[..., None, :, :2])
+    rb = torch.min(boxes1[..., :, None, 2:], boxes2[..., None, :, 2:])
+
+    wh = (rb - lt).clamp(min=0)  # (..., N, M, 2)
+    inter = wh[..., 0] * wh[..., 1]  # (..., N, M)
+
+    union = area1[..., None] + area2[..., None, :] - inter
+
+    iou = inter / union
+    return iou, union
+
+
+def generalized_box_iou(boxes1, boxes2):
+    """
+    Batched version of Generalized IoU from https://giou.stanford.edu/
+
+    Boxes should be in [x0, y0, x1, y1] format
+
+    Inputs:
+    - boxes1: Tensor of shape (..., N, 4)
+    - boxes2: Tensor of shape (..., M, 4)
+
+    Returns:
+    - giou: Tensor of shape (..., N, M)
+    """
+    iou, union = box_iou(boxes1, boxes2)
+
+    # boxes1: (..., N, 4) -> (..., N, 1, 2)
+    # boxes2: (..., M, 4) -> (..., 1, M, 2)
+    lt = torch.min(boxes1[..., :, None, :2], boxes2[..., None, :, :2])
+    rb = torch.max(boxes1[..., :, None, 2:], boxes2[..., None, :, 2:])
+
+    wh = (rb - lt).clamp(min=0)  # (..., N, M, 2)
+    area = wh[..., 0] * wh[..., 1]  # (..., N, M)
+
+    return iou - (area - union) / area
+
+
+@torch.jit.script
+def fast_diag_generalized_box_iou(boxes1, boxes2):
+    assert len(boxes1) == len(boxes2)
+    box1_xy = boxes1[:, 2:]
+    box1_XY = boxes1[:, :2]
+    box2_xy = boxes2[:, 2:]
+    box2_XY = boxes2[:, :2]
+    # assert (box1_xy >= box1_XY).all()
+    # assert (box2_xy >= box2_XY).all()
+    area1 = (box1_xy - box1_XY).prod(-1)
+    area2 = (box2_xy - box2_XY).prod(-1)
+
+    lt = torch.max(box1_XY, box2_XY)  # [N,2]
+    lt2 = torch.min(box1_XY, box2_XY)
+    rb = torch.min(box1_xy, box2_xy)  # [N,2]
+    rb2 = torch.max(box1_xy, box2_xy)
+
+    inter = (rb - lt).clamp(min=0).prod(-1)
+    tot_area = (rb2 - lt2).clamp(min=0).prod(-1)
+
+    union = area1 + area2 - inter
+
+    iou = inter / union
+
+    return iou - (tot_area - union) / tot_area
+
+
+@torch.jit.script
+def fast_diag_box_iou(boxes1, boxes2):
+    assert len(boxes1) == len(boxes2)
+    box1_xy = boxes1[:, 2:]
+    box1_XY = boxes1[:, :2]
+    box2_xy = boxes2[:, 2:]
+    box2_XY = boxes2[:, :2]
+    # assert (box1_xy >= box1_XY).all()
+    # assert (box2_xy >= box2_XY).all()
+    area1 = (box1_xy - box1_XY).prod(-1)
+    area2 = (box2_xy - box2_XY).prod(-1)
+
+    lt = torch.max(box1_XY, box2_XY)  # [N,2]
+    rb = torch.min(box1_xy, box2_xy)  # [N,2]
+
+    inter = (rb - lt).clamp(min=0).prod(-1)
+
+    union = area1 + area2 - inter
+
+    iou = inter / union
+
+    return iou
+
+
+def box_xywh_inter_union(
+    boxes1: torch.Tensor, boxes2: torch.Tensor
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    # Asuumes boxes in xywh format
+    assert boxes1.size(-1) == 4 and boxes2.size(-1) == 4
+    boxes1 = box_xywh_to_xyxy(boxes1)
+    boxes2 = box_xywh_to_xyxy(boxes2)
+    box1_tl_xy = boxes1[..., :2]
+    box1_br_xy = boxes1[..., 2:]
+    box2_tl_xy = boxes2[..., :2]
+    box2_br_xy = boxes2[..., 2:]
+    area1 = (box1_br_xy - box1_tl_xy).prod(-1)
+    area2 = (box2_br_xy - box2_tl_xy).prod(-1)
+
+    assert (area1 >= 0).all() and (area2 >= 0).all()
+    tl = torch.max(box1_tl_xy, box2_tl_xy)
+    br = torch.min(box1_br_xy, box2_br_xy)
+
+    inter = (br - tl).clamp(min=0).prod(-1)
+    union = area1 + area2 - inter
+
+    return inter, union

sam3/model/data_misc.py

@@ -0,0 +1,209 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+"""
+Misc functions, including distributed helpers.
+"""
+
+import collections
+import re
+
+from dataclasses import dataclass, field as field_ptr_behaviour, fields, is_dataclass
+from typing import Any, get_args, get_origin, List, Mapping, Optional, Sequence, Union
+
+import torch
+
+
+MyTensor = Union[torch.Tensor, List[Any]]
+
+
+def interpolate(
+    input, size=None, scale_factor=None, mode="nearest", align_corners=None
+):
+    # type: (Tensor, Optional[List[int]], Optional[float], str, Optional[bool]) -> Tensor
+    """
+    Equivalent to nn.functional.interpolate, but with support for empty channel sizes.
+    """
+    if input.numel() > 0:
+        return torch.nn.functional.interpolate(
+            input, size, scale_factor, mode, align_corners
+        )
+
+    assert (
+        input.shape[0] != 0 or input.shape[1] != 0
+    ), "At least one of the two first dimensions must be non zero"
+
+    if input.shape[1] == 0:
+        # Pytorch doesn't support null dimension on the channel dimension, so we transpose to fake a null batch dim
+        return torch.nn.functional.interpolate(
+            input.transpose(0, 1), size, scale_factor, mode, align_corners
+        ).transpose(0, 1)
+
+    # empty batch dimension is now supported in pytorch
+    return torch.nn.functional.interpolate(
+        input, size, scale_factor, mode, align_corners
+    )
+
+
+@dataclass
+class BatchedPointer:
+    stage_ids: MyTensor
+    stage_ids__type = torch.long
+    query_ids: MyTensor
+    query_ids__type = torch.long
+    object_ids: MyTensor
+    object_ids__type = torch.long
+    ptr_mask: MyTensor
+    ptr_mask__type = torch.bool
+    ptr_types: MyTensor
+    ptr_types__type = torch.long
+
+
+@dataclass
+class FindStage:
+    img_ids: MyTensor
+    img_ids__type = torch.long
+    text_ids: MyTensor
+    text_ids__type = torch.long
+
+    input_boxes: MyTensor
+    input_boxes__type = torch.float
+    input_boxes_mask: MyTensor
+    input_boxes_mask__type = torch.bool
+    input_boxes_label: MyTensor
+    input_boxes_label__type = torch.long
+
+    input_points: MyTensor
+    input_points__type = torch.float
+    input_points_mask: MyTensor
+    input_points_mask__type = torch.bool
+
+    # We track the object ids referred to by this query.
+    # This is beneficial for tracking in videos without the need for pointers.
+    object_ids: Optional[List[List]] = None  # List of objects per query
+
+
+@dataclass
+class BatchedFindTarget:
+    # The number of boxes in each find query
+    num_boxes: MyTensor
+    num_boxes__type = torch.long
+
+    # Target boxes in normalized CxCywh format
+    boxes: MyTensor
+    boxes__type = torch.float
+    # Target boxes in normalized CxCywh format but in padded representation
+    # as used in BinaryHungarianMatcherV2 (unlike the packed ones in `boxes`)
+    boxes_padded: MyTensor
+    boxes_padded__type = torch.float
+
+    # For hybrid matching, we repeat the boxes
+    repeated_boxes: MyTensor
+    repeated_boxes__type = torch.float
+
+    # Target Segmentation masks
+    segments: Optional[MyTensor]
+    segments__type = torch.bool
+
+    # Target Semantic Segmentation masks
+    semantic_segments: Optional[MyTensor]
+    semantic_segments__type = torch.bool
+
+    is_valid_segment: Optional[MyTensor]
+    is_valid_segment__type = torch.bool
+
+    # Whether annotations are exhaustive for each query
+    is_exhaustive: MyTensor
+    is_exhaustive__type = torch.bool
+
+    # The object id for each ground-truth box, in both packed and padded representations
+    object_ids: MyTensor
+    object_ids__type = torch.long
+    object_ids_padded: MyTensor
+    object_ids_padded__type = torch.long
+
+
+@dataclass
+class BatchedInferenceMetadata:
+    """All metadata required to post-process a find stage"""
+
+    # Coco id that corresponds to the "image" for evaluation by the coco evaluator
+    coco_image_id: MyTensor
+    coco_image_id__type = torch.long
+
+    # id in the original dataset, such that we can use the original evaluator
+    original_image_id: MyTensor
+    original_image_id__type = torch.long
+
+    # Original category id (if we want to use the original evaluator)
+    original_category_id: MyTensor
+    original_category_id__type = torch.int
+
+    # Size of the raw image (height, width)
+    original_size: MyTensor
+    original_size__type = torch.long
+
+    # id of the object in the media (track_id for a video)
+    object_id: MyTensor
+    object_id__type = torch.long
+
+    # index of the frame in the media (0 in the case of a single-frame media)
+    frame_index: MyTensor
+    frame_index__type = torch.long
+
+    # Adding for relations inference
+    # get_text_input: List[Optional[str]]
+
+    # Adding for TA conditional inference
+    is_conditioning_only: List[Optional[bool]]
+
+
+@dataclass
+class BatchedDatapoint:
+    img_batch: torch.Tensor
+    find_text_batch: List[str]
+    find_inputs: List[FindStage]
+    find_targets: List[BatchedFindTarget]
+    find_metadatas: List[BatchedInferenceMetadata]
+    raw_images: Optional[List[Any]] = None
+
+
+def convert_my_tensors(obj):
+    def is_optional_field(field) -> bool:
+        return get_origin(field) is Union and type(None) in get_args(field)
+
+    for field in fields(obj):
+        if is_dataclass(getattr(obj, field.name)):
+            convert_my_tensors(getattr(obj, field.name))
+            continue
+
+        field_type = field.type
+        if is_optional_field(field.type):
+            field_type = Union[get_args(field.type)[:-1]]  # Get the Optional field type
+
+        if field_type != MyTensor or getattr(obj, field.name) is None:
+            continue
+
+        elif len(getattr(obj, field.name)) and isinstance(
+            getattr(obj, field.name)[0], torch.Tensor
+        ):
+            stack_dim = 0
+            if field.name in [
+                "input_boxes",
+                "input_boxes_label",
+            ]:
+                stack_dim = 1
+            setattr(
+                obj,
+                field.name,
+                torch.stack(getattr(obj, field.name), dim=stack_dim).to(
+                    getattr(obj, field.name + "__type")
+                ),
+            )
+        else:
+            setattr(
+                obj,
+                field.name,
+                torch.as_tensor(
+                    getattr(obj, field.name), dtype=getattr(obj, field.name + "__type")
+                ),
+            )
+    return obj

sam3/model/decoder.py

@@ -0,0 +1,956 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+"""
+Transformer decoder.
+Inspired from Pytorch's version, adds the pre-norm variant
+"""
+
+from typing import Any, Dict, List, Optional
+
+import numpy as np
+
+import torch
+
+from sam3.sam.transformer import RoPEAttention
+
+from torch import nn, Tensor
+from torchvision.ops.roi_align import RoIAlign
+
+from .act_ckpt_utils import activation_ckpt_wrapper
+
+from .box_ops import box_cxcywh_to_xyxy
+
+from .model_misc import (
+    gen_sineembed_for_position,
+    get_activation_fn,
+    get_clones,
+    inverse_sigmoid,
+    MLP,
+)
+
+
+class TransformerDecoderLayer(nn.Module):
+    def __init__(
+        self,
+        activation: str,
+        d_model: int,
+        dim_feedforward: int,
+        dropout: float,
+        cross_attention: nn.Module,
+        n_heads: int,
+        use_text_cross_attention: bool = False,
+    ):
+        super().__init__()
+
+        # cross attention
+        self.cross_attn = cross_attention
+        self.dropout1 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.norm1 = nn.LayerNorm(d_model)
+
+        # cross attention text
+        self.use_text_cross_attention = use_text_cross_attention
+        if use_text_cross_attention:
+            self.ca_text = nn.MultiheadAttention(d_model, n_heads, dropout=dropout)
+            self.catext_dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+            self.catext_norm = nn.LayerNorm(d_model)
+
+        # self attention
+        self.self_attn = nn.MultiheadAttention(d_model, n_heads, dropout=dropout)
+        self.dropout2 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.norm2 = nn.LayerNorm(d_model)
+
+        # ffn
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.activation = get_activation_fn(activation)
+        self.dropout3 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+        self.dropout4 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.norm3 = nn.LayerNorm(d_model)
+
+    @staticmethod
+    def with_pos_embed(tensor, pos):
+        return tensor if pos is None else tensor + pos
+
+    def forward_ffn(self, tgt):
+        with torch.amp.autocast(device_type="cuda", enabled=False):
+            tgt2 = self.linear2(self.dropout3(self.activation(self.linear1(tgt))))
+        tgt = tgt + self.dropout4(tgt2)
+        tgt = self.norm3(tgt)
+        return tgt
+
+    def forward(
+        self,
+        # for tgt
+        tgt: Optional[Tensor],  # nq, bs, d_model
+        tgt_query_pos: Optional[Tensor] = None,  # pos for query. MLP(Sine(pos))
+        tgt_query_sine_embed: Optional[Tensor] = None,  # pos for query. Sine(pos)
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        tgt_reference_points: Optional[Tensor] = None,  # nq, bs, 4
+        memory_text: Optional[Tensor] = None,  # num_token, bs, d_model
+        text_attention_mask: Optional[Tensor] = None,  # bs, num_token
+        # for memory
+        memory: Optional[Tensor] = None,  # hw, bs, d_model
+        memory_key_padding_mask: Optional[Tensor] = None,
+        memory_level_start_index: Optional[Tensor] = None,  # num_levels
+        memory_spatial_shapes: Optional[Tensor] = None,  # bs, num_levels, 2
+        memory_pos: Optional[Tensor] = None,  # pos for memory
+        # sa
+        self_attn_mask: Optional[Tensor] = None,  # mask used for self-attention
+        cross_attn_mask: Optional[Tensor] = None,  # mask used for cross-attention
+        # dac
+        dac=False,
+        dac_use_selfatt_ln=True,
+        presence_token=None,
+        # skip inside deformable attn
+        identity=0.0,
+        **kwargs,  # additional kwargs for compatibility
+    ):
+        """
+        Input:
+            - tgt/tgt_query_pos: nq, bs, d_model
+            -
+        """
+        # self attention
+        if self.self_attn is not None:
+            if dac:
+                # we only apply self attention to the first half of the queries
+                assert tgt.shape[0] % 2 == 0
+                num_o2o_queries = tgt.shape[0] // 2
+                tgt_o2o = tgt[:num_o2o_queries]
+                tgt_query_pos_o2o = tgt_query_pos[:num_o2o_queries]
+                tgt_o2m = tgt[num_o2o_queries:]
+            else:
+                tgt_o2o = tgt
+                tgt_query_pos_o2o = tgt_query_pos
+
+            if presence_token is not None:
+                tgt_o2o = torch.cat([presence_token, tgt_o2o], dim=0)
+                tgt_query_pos_o2o = torch.cat(
+                    [torch.zeros_like(presence_token), tgt_query_pos_o2o], dim=0
+                )
+                tgt_query_pos = torch.cat(
+                    [torch.zeros_like(presence_token), tgt_query_pos], dim=0
+                )
+
+            q = k = self.with_pos_embed(tgt_o2o, tgt_query_pos_o2o)
+            tgt2 = self.self_attn(q, k, tgt_o2o, attn_mask=self_attn_mask)[0]
+            tgt_o2o = tgt_o2o + self.dropout2(tgt2)
+            if dac:
+                if not dac_use_selfatt_ln:
+                    tgt_o2o = self.norm2(tgt_o2o)
+                tgt = torch.cat((tgt_o2o, tgt_o2m), dim=0)  # Recombine
+                if dac_use_selfatt_ln:
+                    tgt = self.norm2(tgt)
+            else:
+                tgt = tgt_o2o
+                tgt = self.norm2(tgt)
+
+        if self.use_text_cross_attention:
+            tgt2 = self.ca_text(
+                self.with_pos_embed(tgt, tgt_query_pos),
+                memory_text,
+                memory_text,
+                key_padding_mask=text_attention_mask,
+            )[0]
+            tgt = tgt + self.catext_dropout(tgt2)
+            tgt = self.catext_norm(tgt)
+
+        if presence_token is not None:
+            presence_token_mask = torch.zeros_like(cross_attn_mask[:, :1, :])
+            cross_attn_mask = torch.cat(
+                [presence_token_mask, cross_attn_mask], dim=1
+            )  # (bs*nheads, 1+nq, hw)
+
+        # Cross attention to image
+        tgt2 = self.cross_attn(
+            query=self.with_pos_embed(tgt, tgt_query_pos),
+            key=self.with_pos_embed(memory, memory_pos),
+            value=memory,
+            attn_mask=cross_attn_mask,
+            key_padding_mask=(
+                memory_key_padding_mask.transpose(0, 1)
+                if memory_key_padding_mask is not None
+                else None
+            ),
+        )[0]
+
+        tgt = tgt + self.dropout1(tgt2)
+        tgt = self.norm1(tgt)
+
+        # ffn
+        tgt = self.forward_ffn(tgt)
+
+        presence_token_out = None
+        if presence_token is not None:
+            presence_token_out = tgt[:1]
+            tgt = tgt[1:]
+
+        return tgt, presence_token_out
+
+
+class TransformerDecoder(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        frozen: bool,
+        interaction_layer,
+        layer,
+        num_layers: int,
+        num_queries: int,
+        return_intermediate: bool,
+        box_refine: bool = False,
+        num_o2m_queries: int = 0,
+        dac: bool = False,
+        boxRPB: str = "none",
+        # Experimental: An object query for SAM 2 tasks
+        instance_query: bool = False,
+        # Defines the number of additional instance queries,
+        # 1 or 4 are the most likely for single vs multi mask support
+        num_instances: int = 1,  # Irrelevant if instance_query is False
+        dac_use_selfatt_ln: bool = True,
+        use_act_checkpoint: bool = False,
+        compile_mode=None,
+        presence_token: bool = False,
+        clamp_presence_logits: bool = True,
+        clamp_presence_logit_max_val: float = 10.0,
+        use_normed_output_consistently: bool = True,
+        separate_box_head_instance: bool = False,
+        separate_norm_instance: bool = False,
+        resolution: Optional[int] = None,
+        stride: Optional[int] = None,
+    ):
+        super().__init__()
+        self.d_model = d_model
+        self.layers = get_clones(layer, num_layers)
+        self.fine_layers = (
+            get_clones(interaction_layer, num_layers)
+            if interaction_layer is not None
+            else [None] * num_layers
+        )
+        self.num_layers = num_layers
+        self.num_queries = num_queries
+        self.dac = dac
+        if dac:
+            self.num_o2m_queries = num_queries
+            tot_num_queries = num_queries
+        else:
+            self.num_o2m_queries = num_o2m_queries
+            tot_num_queries = num_queries + num_o2m_queries
+        self.norm = nn.LayerNorm(d_model)
+        self.return_intermediate = return_intermediate
+        self.bbox_embed = MLP(d_model, d_model, 4, 3)
+        self.query_embed = nn.Embedding(tot_num_queries, d_model)
+        self.instance_query_embed = None
+        self.instance_query_reference_points = None
+        self.use_instance_query = instance_query
+        self.num_instances = num_instances
+        self.use_normed_output_consistently = use_normed_output_consistently
+
+        self.instance_norm = nn.LayerNorm(d_model) if separate_norm_instance else None
+        self.instance_bbox_embed = None
+        if separate_box_head_instance:
+            self.instance_bbox_embed = MLP(d_model, d_model, 4, 3)
+        if instance_query:
+            self.instance_query_embed = nn.Embedding(num_instances, d_model)
+        self.box_refine = box_refine
+        if box_refine:
+            nn.init.constant_(self.bbox_embed.layers[-1].weight.data, 0)
+            nn.init.constant_(self.bbox_embed.layers[-1].bias.data, 0)
+
+            self.reference_points = nn.Embedding(num_queries, 4)
+            if instance_query:
+                self.instance_reference_points = nn.Embedding(num_instances, 4)
+
+        assert boxRPB in ["none", "log", "linear", "both"]
+        self.boxRPB = boxRPB
+        if boxRPB != "none":
+            try:
+                nheads = self.layers[0].cross_attn_image.num_heads
+            except AttributeError:
+                nheads = self.layers[0].cross_attn.num_heads
+
+            n_input = 4 if boxRPB == "both" else 2
+            self.boxRPB_embed_x = MLP(n_input, d_model, nheads, 2)
+            self.boxRPB_embed_y = MLP(n_input, d_model, nheads, 2)
+            self.compilable_cord_cache = None
+            self.compilable_stored_size = None
+            self.coord_cache = {}
+
+            if resolution is not None and stride is not None:
+                feat_size = resolution // stride
+                coords_h, coords_w = self._get_coords(
+                    feat_size, feat_size, device="cuda"
+                )
+                self.compilable_cord_cache = (coords_h, coords_w)
+                self.compilable_stored_size = (feat_size, feat_size)
+
+        self.roi_pooler = (
+            RoIAlign(output_size=7, spatial_scale=1, sampling_ratio=-1, aligned=True)
+            if interaction_layer is not None
+            else None
+        )
+        if frozen:
+            for p in self.parameters():
+                p.requires_grad_(False)
+
+        self.presence_token = None
+        self.clamp_presence_logits = clamp_presence_logits
+        self.clamp_presence_logit_max_val = clamp_presence_logit_max_val
+        if presence_token:
+            self.presence_token = nn.Embedding(1, d_model)
+            self.presence_token_head = MLP(d_model, d_model, 1, 3)
+            self.presence_token_out_norm = nn.LayerNorm(d_model)
+
+        self.ref_point_head = MLP(2 * self.d_model, self.d_model, self.d_model, 2)
+        self.dac_use_selfatt_ln = dac_use_selfatt_ln
+        self.use_act_checkpoint = use_act_checkpoint
+
+        nn.init.normal_(self.query_embed.weight.data)
+        if self.instance_query_embed is not None:
+            nn.init.normal_(self.instance_query_embed.weight.data)
+
+        assert self.roi_pooler is None
+        assert self.return_intermediate, "support return_intermediate only"
+        assert self.box_refine, "support box refine only"
+
+        self.compile_mode = compile_mode
+        self.compiled = False
+        # We defer compilation till after the first forward, to first warm-up the boxRPB cache
+
+        # assign layer index to each layer so that some layers can decide what to do
+        # based on which layer index they are (e.g. cross attention to memory bank only
+        # in selected layers)
+        for layer_idx, layer in enumerate(self.layers):
+            layer.layer_idx = layer_idx
+
+    @staticmethod
+    def _get_coords(H, W, device):
+        coords_h = torch.arange(0, H, device=device, dtype=torch.float32) / H
+        coords_w = torch.arange(0, W, device=device, dtype=torch.float32) / W
+        return coords_h, coords_w
+
+    def _get_rpb_matrix(self, reference_boxes, feat_size):
+        H, W = feat_size
+        boxes_xyxy = box_cxcywh_to_xyxy(reference_boxes).transpose(0, 1)
+        bs, num_queries, _ = boxes_xyxy.shape
+        if self.compilable_cord_cache is None:
+            self.compilable_cord_cache = self._get_coords(H, W, reference_boxes.device)
+            self.compilable_stored_size = (H, W)
+
+        if torch.compiler.is_dynamo_compiling() or self.compilable_stored_size == (
+            H,
+            W,
+        ):
+            # good, hitting the cache, will be compilable
+            coords_h, coords_w = self.compilable_cord_cache
+        else:
+            # cache miss, will create compilation issue
+            # In case we're not compiling, we'll still rely on the dict-based cache
+            if feat_size not in self.coord_cache:
+                self.coord_cache[feat_size] = self._get_coords(
+                    H, W, reference_boxes.device
+                )
+            coords_h, coords_w = self.coord_cache[feat_size]
+
+            assert coords_h.shape == (H,)
+            assert coords_w.shape == (W,)
+
+        deltas_y = coords_h.view(1, -1, 1) - boxes_xyxy.reshape(-1, 1, 4)[:, :, 1:4:2]
+        deltas_y = deltas_y.view(bs, num_queries, -1, 2)
+        deltas_x = coords_w.view(1, -1, 1) - boxes_xyxy.reshape(-1, 1, 4)[:, :, 0:3:2]
+        deltas_x = deltas_x.view(bs, num_queries, -1, 2)
+
+        if self.boxRPB in ["log", "both"]:
+            deltas_x_log = deltas_x * 8  # normalize to -8, 8
+            deltas_x_log = (
+                torch.sign(deltas_x_log)
+                * torch.log2(torch.abs(deltas_x_log) + 1.0)
+                / np.log2(8)
+            )
+
+            deltas_y_log = deltas_y * 8  # normalize to -8, 8
+            deltas_y_log = (
+                torch.sign(deltas_y_log)
+                * torch.log2(torch.abs(deltas_y_log) + 1.0)
+                / np.log2(8)
+            )
+            if self.boxRPB == "log":
+                deltas_x = deltas_x_log
+                deltas_y = deltas_y_log
+            else:
+                deltas_x = torch.cat([deltas_x, deltas_x_log], dim=-1)
+                deltas_y = torch.cat([deltas_y, deltas_y_log], dim=-1)
+
+        if self.training:
+            assert self.use_act_checkpoint, "activation ckpt not enabled in decoder"
+        deltas_x = activation_ckpt_wrapper(self.boxRPB_embed_x)(
+            x=deltas_x,
+            act_ckpt_enable=self.training and self.use_act_checkpoint,
+        )  # bs, num_queries, W, n_heads
+        deltas_y = activation_ckpt_wrapper(self.boxRPB_embed_y)(
+            x=deltas_y,
+            act_ckpt_enable=self.training and self.use_act_checkpoint,
+        )  # bs, num_queries, H, n_heads
+
+        if not torch.compiler.is_dynamo_compiling():
+            assert deltas_x.shape[:3] == (bs, num_queries, W)
+            assert deltas_y.shape[:3] == (bs, num_queries, H)
+
+        B = deltas_y.unsqueeze(3) + deltas_x.unsqueeze(
+            2
+        )  # bs, num_queries, H, W, n_heads
+        if not torch.compiler.is_dynamo_compiling():
+            assert B.shape[:4] == (bs, num_queries, H, W)
+        B = B.flatten(2, 3)  # bs, num_queries, H*W, n_heads
+        B = B.permute(0, 3, 1, 2)  # bs, n_heads, num_queries, H*W
+        B = B.contiguous()  # memeff attn likes ordered strides
+        if not torch.compiler.is_dynamo_compiling():
+            assert B.shape[2:] == (num_queries, H * W)
+        return B
+
+    def forward(
+        self,
+        tgt,
+        memory,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        reference_boxes: Optional[Tensor] = None,  # num_queries, bs, 4
+        # for memory
+        level_start_index: Optional[Tensor] = None,  # num_levels
+        spatial_shapes: Optional[Tensor] = None,  # bs, num_levels, 2
+        valid_ratios: Optional[Tensor] = None,
+        # for text
+        memory_text: Optional[Tensor] = None,
+        text_attention_mask: Optional[Tensor] = None,
+        # if `apply_dac` is None, it will default to `self.dac`
+        apply_dac: Optional[bool] = None,
+        is_instance_prompt=False,
+        decoder_extra_kwargs: Optional[Dict] = None,
+        # ROI memory bank
+        obj_roi_memory_feat=None,
+        obj_roi_memory_mask=None,
+        box_head_trk=None,
+    ):
+        """
+        Input:
+            - tgt: nq, bs, d_model
+            - memory: \\sum{hw}, bs, d_model
+            - pos: \\sum{hw}, bs, d_model
+            - reference_boxes: nq, bs, 4 (after sigmoid)
+            - valid_ratios/spatial_shapes: bs, nlevel, 2
+        """
+        if memory_mask is not None:
+            assert (
+                self.boxRPB == "none"
+            ), "inputting a memory_mask in the presence of boxRPB is unexpected/not implemented"
+
+        apply_dac = apply_dac if apply_dac is not None else self.dac
+        if apply_dac:
+            assert (tgt.shape[0] == self.num_queries) or (
+                self.use_instance_query
+                and (tgt.shape[0] == self.instance_query_embed.num_embeddings)
+            )
+
+            tgt = tgt.repeat(2, 1, 1)
+            # note that we don't tile tgt_mask, since DAC doesn't
+            # use self-attention in o2m queries
+            if reference_boxes is not None:
+                assert (reference_boxes.shape[0] == self.num_queries) or (
+                    self.use_instance_query
+                    and (
+                        reference_boxes.shape[0]
+                        == self.instance_query_embed.num_embeddings
+                    )
+                )
+                reference_boxes = reference_boxes.repeat(2, 1, 1)
+
+        bs = tgt.shape[1]
+        intermediate = []
+        intermediate_presence_logits = []
+        presence_feats = None
+
+        if self.box_refine:
+            if reference_boxes is None:
+                # In this case, we're in a one-stage model, so we generate the reference boxes
+                reference_boxes = self.reference_points.weight.unsqueeze(1)
+                reference_boxes = (
+                    reference_boxes.repeat(2, bs, 1)
+                    if apply_dac
+                    else reference_boxes.repeat(1, bs, 1)
+                )
+                reference_boxes = reference_boxes.sigmoid()
+            intermediate_ref_boxes = [reference_boxes]
+        else:
+            reference_boxes = None
+            intermediate_ref_boxes = None
+
+        output = tgt
+        presence_out = None
+        if self.presence_token is not None and is_instance_prompt is False:
+            # expand to batch dim
+            presence_out = self.presence_token.weight[None].expand(1, bs, -1)
+
+        box_head = self.bbox_embed
+        if is_instance_prompt and self.instance_bbox_embed is not None:
+            box_head = self.instance_bbox_embed
+
+        out_norm = self.norm
+        if is_instance_prompt and self.instance_norm is not None:
+            out_norm = self.instance_norm
+
+        for layer_idx, layer in enumerate(self.layers):
+            reference_points_input = (
+                reference_boxes[:, :, None]
+                * torch.cat([valid_ratios, valid_ratios], -1)[None, :]
+            )  # nq, bs, nlevel, 4
+
+            query_sine_embed = gen_sineembed_for_position(
+                reference_points_input[:, :, 0, :], self.d_model
+            )  # nq, bs, d_model*2
+
+            # conditional query
+            query_pos = self.ref_point_head(query_sine_embed)  # nq, bs, d_model
+
+            if self.boxRPB != "none" and reference_boxes is not None:
+                assert (
+                    spatial_shapes.shape[0] == 1
+                ), "only single scale support implemented"
+                memory_mask = self._get_rpb_matrix(
+                    reference_boxes,
+                    (spatial_shapes[0, 0], spatial_shapes[0, 1]),
+                )
+                memory_mask = memory_mask.flatten(0, 1)  # (bs*n_heads, nq, H*W)
+            if self.training:
+                assert (
+                    self.use_act_checkpoint
+                ), "Activation checkpointing not enabled in the decoder"
+            output, presence_out = activation_ckpt_wrapper(layer)(
+                tgt=output,
+                tgt_query_pos=query_pos,
+                tgt_query_sine_embed=query_sine_embed,
+                tgt_key_padding_mask=tgt_key_padding_mask,
+                tgt_reference_points=reference_points_input,
+                memory_text=memory_text,
+                text_attention_mask=text_attention_mask,
+                memory=memory,
+                memory_key_padding_mask=memory_key_padding_mask,
+                memory_level_start_index=level_start_index,
+                memory_spatial_shapes=spatial_shapes,
+                memory_pos=pos,
+                self_attn_mask=tgt_mask,
+                cross_attn_mask=memory_mask,
+                dac=apply_dac,
+                dac_use_selfatt_ln=self.dac_use_selfatt_ln,
+                presence_token=presence_out,
+                **(decoder_extra_kwargs or {}),
+                act_ckpt_enable=self.training and self.use_act_checkpoint,
+                # ROI memory bank
+                obj_roi_memory_feat=obj_roi_memory_feat,
+                obj_roi_memory_mask=obj_roi_memory_mask,
+            )
+
+            # iter update
+            if self.box_refine:
+                reference_before_sigmoid = inverse_sigmoid(reference_boxes)
+                if box_head_trk is None:
+                    # delta_unsig = self.bbox_embed(output)
+                    if not self.use_normed_output_consistently:
+                        delta_unsig = box_head(output)
+                    else:
+                        delta_unsig = box_head(out_norm(output))
+                else:
+                    # box_head_trk use a separate box head for tracking queries
+                    Q_det = decoder_extra_kwargs["Q_det"]
+                    assert output.size(0) >= Q_det
+                    delta_unsig_det = self.bbox_embed(output[:Q_det])
+                    delta_unsig_trk = box_head_trk(output[Q_det:])
+                    delta_unsig = torch.cat([delta_unsig_det, delta_unsig_trk], dim=0)
+                outputs_unsig = delta_unsig + reference_before_sigmoid
+                new_reference_points = outputs_unsig.sigmoid()
+
+                reference_boxes = new_reference_points.detach()
+                if layer_idx != self.num_layers - 1:
+                    intermediate_ref_boxes.append(new_reference_points)
+            else:
+                raise NotImplementedError("not implemented yet")
+
+            intermediate.append(out_norm(output))
+            if self.presence_token is not None and is_instance_prompt is False:
+                # norm, mlp head
+                intermediate_layer_presence_logits = self.presence_token_head(
+                    self.presence_token_out_norm(presence_out)
+                ).squeeze(-1)
+
+                # clamp to mitigate numerical issues
+                if self.clamp_presence_logits:
+                    intermediate_layer_presence_logits.clamp(
+                        min=-self.clamp_presence_logit_max_val,
+                        max=self.clamp_presence_logit_max_val,
+                    )
+
+                intermediate_presence_logits.append(intermediate_layer_presence_logits)
+                presence_feats = presence_out.clone()
+
+        if not self.compiled and self.compile_mode is not None:
+            self.forward = torch.compile(
+                self.forward, mode=self.compile_mode, fullgraph=True
+            )
+            self.compiled = True
+
+        return (
+            torch.stack(intermediate),
+            torch.stack(intermediate_ref_boxes),
+            (
+                torch.stack(intermediate_presence_logits)
+                if self.presence_token is not None and is_instance_prompt is False
+                else None
+            ),
+            presence_feats,
+        )
+
+
+class TransformerEncoderCrossAttention(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        frozen: bool,
+        pos_enc_at_input: bool,
+        layer,
+        num_layers: int,
+        use_act_checkpoint: bool = False,
+        batch_first: bool = False,  # Do layers expect batch first input?
+        # which layers to exclude cross attention? default: None, means all
+        # layers use cross attention
+        remove_cross_attention_layers: Optional[list] = None,
+    ):
+        super().__init__()
+        self.d_model = d_model
+        self.layers = get_clones(layer, num_layers)
+        self.num_layers = num_layers
+        self.norm = nn.LayerNorm(d_model)
+        self.pos_enc_at_input = pos_enc_at_input
+        self.use_act_checkpoint = use_act_checkpoint
+
+        if frozen:
+            for p in self.parameters():
+                p.requires_grad_(False)
+
+        self.batch_first = batch_first
+
+        # remove cross attention layers if specified
+        self.remove_cross_attention_layers = [False] * self.num_layers
+        if remove_cross_attention_layers is not None:
+            for i in remove_cross_attention_layers:
+                self.remove_cross_attention_layers[i] = True
+        assert len(self.remove_cross_attention_layers) == len(self.layers)
+
+        for i, remove_cross_attention in enumerate(self.remove_cross_attention_layers):
+            if remove_cross_attention:
+                self.layers[i].cross_attn_image = None
+                self.layers[i].norm2 = None
+                self.layers[i].dropout2 = None
+
+    def forward(
+        self,
+        src,  # self-attention inputs
+        prompt,  # cross-attention inputs
+        src_mask: Optional[Tensor] = None,  # att.mask for self-attention inputs
+        prompt_mask: Optional[Tensor] = None,  # att.mask for cross-attention inputs
+        src_key_padding_mask: Optional[Tensor] = None,
+        prompt_key_padding_mask: Optional[Tensor] = None,
+        src_pos: Optional[Tensor] = None,  # pos_enc for self-attention inputs
+        prompt_pos: Optional[Tensor] = None,  # pos_enc for cross-attention inputs
+        feat_sizes: Optional[list] = None,
+        num_obj_ptr_tokens: int = 0,  # number of object pointer *tokens*
+    ):
+        if isinstance(src, list):
+            assert isinstance(src_key_padding_mask, list) and isinstance(src_pos, list)
+            assert len(src) == len(src_key_padding_mask) == len(src_pos) == 1
+            src, src_key_padding_mask, src_pos = (
+                src[0],
+                src_key_padding_mask[0],
+                src_pos[0],
+            )
+
+        assert (
+            src.shape[1] == prompt.shape[1]
+        ), "Batch size must be the same for src and prompt"
+
+        output = src
+
+        if self.pos_enc_at_input and src_pos is not None:
+            output = output + 0.1 * src_pos
+
+        if self.batch_first:
+            # Convert to batch first
+            output = output.transpose(0, 1)
+            src_pos = src_pos.transpose(0, 1)
+            prompt = prompt.transpose(0, 1)
+            prompt_pos = prompt_pos.transpose(0, 1)
+
+        for layer in self.layers:
+            kwds = {}
+            if isinstance(layer.cross_attn_image, RoPEAttention):
+                kwds = {"num_k_exclude_rope": num_obj_ptr_tokens}
+
+            output = activation_ckpt_wrapper(layer)(
+                tgt=output,
+                memory=prompt,
+                tgt_mask=src_mask,
+                memory_mask=prompt_mask,
+                tgt_key_padding_mask=src_key_padding_mask,
+                memory_key_padding_mask=prompt_key_padding_mask,
+                pos=prompt_pos,
+                query_pos=src_pos,
+                dac=False,
+                attn_bias=None,
+                act_ckpt_enable=self.training and self.use_act_checkpoint,
+                **kwds,
+            )
+            normed_output = self.norm(output)
+
+        if self.batch_first:
+            # Convert back to seq first
+            normed_output = normed_output.transpose(0, 1)
+            src_pos = src_pos.transpose(0, 1)
+
+        return {
+            "memory": normed_output,
+            "pos_embed": src_pos,
+            "padding_mask": src_key_padding_mask,
+        }
+
+
+class TransformerDecoderLayerv1(nn.Module):
+    def __init__(
+        self,
+        activation: str,
+        cross_attention: nn.Module,
+        d_model: int,
+        dim_feedforward: int,
+        dropout: float,
+        pos_enc_at_attn: bool,
+        pos_enc_at_cross_attn_keys: bool,
+        pos_enc_at_cross_attn_queries: bool,
+        pre_norm: bool,
+        self_attention: nn.Module,
+    ):
+        super().__init__()
+        self.d_model = d_model
+        self.dim_feedforward = dim_feedforward
+        self.dropout_value = dropout
+        self.self_attn = self_attention
+        self.cross_attn_image = cross_attention
+
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.activation_str = activation
+        self.activation = get_activation_fn(activation)
+        self.pre_norm = pre_norm
+
+        self.pos_enc_at_attn = pos_enc_at_attn
+        self.pos_enc_at_cross_attn_queries = pos_enc_at_cross_attn_queries
+        self.pos_enc_at_cross_attn_keys = pos_enc_at_cross_attn_keys
+
+    def forward_post(
+        self,
+        tgt,
+        memory,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        **kwargs,
+    ):
+        q = k = tgt + query_pos if self.pos_enc_at_attn else tgt
+
+        # Self attention
+        tgt2 = self.self_attn(
+            q,
+            k,
+            value=tgt,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout1(tgt2)
+        tgt = self.norm1(tgt)
+
+        # Cross attention to image
+        tgt2 = self.cross_attn_image(
+            query=tgt + query_pos if self.pos_enc_at_cross_attn_queries else tgt,
+            key=memory + pos if self.pos_enc_at_cross_attn_keys else memory,
+            value=memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout2(tgt2)
+        tgt = self.norm2(tgt)
+
+        # FFN
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
+        tgt = tgt + self.dropout3(tgt2)
+        tgt = self.norm3(tgt)
+        return tgt
+
+    def forward_pre(
+        self,
+        tgt,
+        memory,
+        dac: bool = False,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        attn_bias: Optional[Tensor] = None,
+        **kwargs,
+    ):
+        if dac:
+            # we only apply self attention to the first half of the queries
+            assert tgt.shape[0] % 2 == 0
+            other_tgt = tgt[tgt.shape[0] // 2 :]
+            tgt = tgt[: tgt.shape[0] // 2]
+        tgt2 = self.norm1(tgt)
+        q = k = tgt2 + query_pos if self.pos_enc_at_attn else tgt2
+        tgt2 = self.self_attn(
+            q,
+            k,
+            value=tgt2,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout1(tgt2)
+        if dac:
+            # Recombine
+            tgt = torch.cat((tgt, other_tgt), dim=0)
+        tgt2 = self.norm2(tgt)
+        tgt2 = self.cross_attn_image(
+            query=tgt2 + query_pos if self.pos_enc_at_cross_attn_queries else tgt2,
+            key=memory + pos if self.pos_enc_at_cross_attn_keys else memory,
+            value=memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+            attn_bias=attn_bias,
+        )[0]
+        tgt = tgt + self.dropout2(tgt2)
+        tgt2 = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
+        tgt = tgt + self.dropout3(tgt2)
+        return tgt
+
+    def forward(
+        self,
+        tgt,
+        memory,
+        dac: bool = False,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        attn_bias: Optional[Tensor] = None,
+        **kwds: Any,
+    ) -> torch.Tensor:
+        fwd_fn = self.forward_pre if self.pre_norm else self.forward_post
+        return fwd_fn(
+            tgt,
+            memory,
+            dac=dac,
+            tgt_mask=tgt_mask,
+            memory_mask=memory_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+            pos=pos,
+            query_pos=query_pos,
+            attn_bias=attn_bias,
+            **kwds,
+        )
+
+
+class TransformerDecoderLayerv2(TransformerDecoderLayerv1):
+    def __init__(self, cross_attention_first=False, *args: Any, **kwds: Any):
+        super().__init__(*args, **kwds)
+        self.cross_attention_first = cross_attention_first
+
+    def _forward_sa(self, tgt, query_pos):
+        # Self-Attention
+        tgt2 = self.norm1(tgt)
+        q = k = tgt2 + query_pos if self.pos_enc_at_attn else tgt2
+        tgt2 = self.self_attn(q, k, v=tgt2)
+        tgt = tgt + self.dropout1(tgt2)
+        return tgt
+
+    def _forward_ca(self, tgt, memory, query_pos, pos, num_k_exclude_rope=0):
+        if self.cross_attn_image is None:
+            return tgt
+
+        kwds = {}
+        if num_k_exclude_rope > 0:
+            assert isinstance(self.cross_attn_image, RoPEAttention)
+            kwds = {"num_k_exclude_rope": num_k_exclude_rope}
+
+        # Cross-Attention
+        tgt2 = self.norm2(tgt)
+        tgt2 = self.cross_attn_image(
+            q=tgt2 + query_pos if self.pos_enc_at_cross_attn_queries else tgt2,
+            k=memory + pos if self.pos_enc_at_cross_attn_keys else memory,
+            v=memory,
+            **kwds,
+        )
+        tgt = tgt + self.dropout2(tgt2)
+        return tgt
+
+    def forward_pre(
+        self,
+        tgt,
+        memory,
+        dac: bool,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        attn_bias: Optional[Tensor] = None,
+        num_k_exclude_rope: int = 0,
+    ):
+        assert dac is False
+        assert tgt_mask is None
+        assert memory_mask is None
+        assert tgt_key_padding_mask is None
+        assert memory_key_padding_mask is None
+        assert attn_bias is None
+
+        if self.cross_attention_first:
+            tgt = self._forward_ca(tgt, memory, query_pos, pos, num_k_exclude_rope)
+            tgt = self._forward_sa(tgt, query_pos)
+        else:
+            tgt = self._forward_sa(tgt, query_pos)
+            tgt = self._forward_ca(tgt, memory, query_pos, pos, num_k_exclude_rope)
+
+        # MLP
+        tgt2 = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
+        tgt = tgt + self.dropout3(tgt2)
+        return tgt
+
+    def forward(self, *args: Any, **kwds: Any) -> torch.Tensor:
+        if self.pre_norm:
+            return self.forward_pre(*args, **kwds)
+        raise NotImplementedError

sam3/model/edt.py

@@ -0,0 +1,173 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+"""Triton kernel for euclidean distance transform (EDT)"""
+
+import torch
+import triton
+import triton.language as tl
+
+"""
+Disclaimer: This implementation is not meant to be extremely efficient. A CUDA kernel would likely be more efficient.
+Even in Triton, there may be more suitable algorithms.
+
+The goal of this kernel is to mimic cv2.distanceTransform(input, cv2.DIST_L2, 0).
+Recall that the euclidean distance transform (EDT) calculates the L2 distance to the closest zero pixel for each pixel of the source image.
+
+For images of size NxN, the naive algorithm would be to compute pairwise distances between every pair of points, leading to a O(N^4) algorithm, which is obviously impractical.
+One can do better using the following approach:
+- First, compute the distance to the closest point in the same row. We can write it as Row_EDT[i,j] = min_k (sqrt((k-j)^2) if input[i,k]==0 else +infinity). With a naive implementation, this step has a O(N^3) complexity
+- Then, because of triangular inequality, we notice that the EDT for a given location [i,j] is the min of the row EDTs in the same column. EDT[i,j] = min_k Row_EDT[k, j]. This is also O(N^3)
+
+Overall, this algorithm is quite amenable to parallelization, and has a complexity O(N^3). Can we do better?
+
+It turns out that we can leverage the structure of the L2 distance (nice and convex) to find the minimum in a more efficient way.
+We follow the algorithm from "Distance Transforms of Sampled Functions" (https://cs.brown.edu/people/pfelzens/papers/dt-final.pdf), which is also what's implemented in opencv
+
+For a single dimension EDT, we can compute the EDT of an arbitrary function F, that we discretize over the grid. Note that for the binary EDT that we're interested in, we can set F(i,j) = 0 if input[i,j]==0 else +infinity
+For now, we'll compute the EDT squared, and will take the sqrt only at the very end.
+The basic idea is that each point at location i spawns a parabola around itself, with a bias equal to F(i). So specifically, we're looking at the parabola (x - i)^2 + F(i)
+When we're looking for the row EDT at location j, we're effectively looking for min_i (x-i)^2 + F(i). In other word we want to find the lowest parabola at location j.
+
+To do this efficiently, we need to maintain the lower envelope of the union of parabolas. This can be constructed on the fly using a sort of stack approach:
+ - every time we want to add a new parabola, we check if it may be covering the current right-most parabola. If so, then that parabola was useless, so we can pop it from the stack
+ - repeat until we can't find any more parabola to pop. Then push the new one.
+
+This algorithm runs in O(N) for a single row, so overall O(N^2) when applied to all rows
+Similarly as before, we notice that we can decompose the algorithm for rows and columns, leading to an overall run-time of O(N^2)
+
+This algorithm is less suited for to GPUs, since the one-dimensional EDT computation is quite sequential in nature. However, we can parallelize over batch and row dimensions.
+In Triton, things are particularly bad at the moment, since there is no support for reading/writing to the local memory at a specific index (a local gather is coming soon, see https://github.com/triton-lang/triton/issues/974, but no mention of writing, ie scatter)
+One could emulate these operations with masking, but in initial tests, it proved to be worst than naively reading and writing to the global memory. My guess is that the cache is compensating somewhat for the repeated single-point accesses.
+
+
+The timing obtained on a H100 for a random batch of masks of dimension 256 x 1024 x 1024 are as follows:
+- OpenCV: 1780ms (including round-trip to cpu, but discounting the fact that it introduces a synchronization point)
+- triton, O(N^3) algo: 627ms
+- triton, O(N^2) algo: 322ms
+
+Overall, despite being quite naive, this implementation is roughly 5.5x faster than the openCV cpu implem
+
+"""
+
+
+@triton.jit
+def edt_kernel(inputs_ptr, outputs_ptr, v, z, height, width, horizontal: tl.constexpr):
+    # This is a somewhat verbatim implementation of the efficient 1D EDT algorithm described above
+    # It can be applied horizontally or vertically depending if we're doing the first or second stage.
+    # It's parallelized across batch+row (or batch+col if horizontal=False)
+    # TODO: perhaps the implementation can be revisited if/when local gather/scatter become available in triton
+    batch_id = tl.program_id(axis=0)
+    if horizontal:
+        row_id = tl.program_id(axis=1)
+        block_start = (batch_id * height * width) + row_id * width
+        length = width
+        stride = 1
+    else:
+        col_id = tl.program_id(axis=1)
+        block_start = (batch_id * height * width) + col_id
+        length = height
+        stride = width
+
+    # This will be the index of the right most parabola in the envelope ("the top of the stack")
+    k = 0
+    for q in range(1, length):
+        # Read the function value at the current location. Note that we're doing a singular read, not very efficient
+        cur_input = tl.load(inputs_ptr + block_start + (q * stride))
+        # location of the parabola on top of the stack
+        r = tl.load(v + block_start + (k * stride))
+        # associated boundary
+        z_k = tl.load(z + block_start + (k * stride))
+        # value of the function at the parabola location
+        previous_input = tl.load(inputs_ptr + block_start + (r * stride))
+        # intersection between the two parabolas
+        s = (cur_input - previous_input + q * q - r * r) / (q - r) / 2
+
+        # we'll pop as many parabolas as required
+        while s <= z_k and k - 1 >= 0:
+            k = k - 1
+            r = tl.load(v + block_start + (k * stride))
+            z_k = tl.load(z + block_start + (k * stride))
+            previous_input = tl.load(inputs_ptr + block_start + (r * stride))
+            s = (cur_input - previous_input + q * q - r * r) / (q - r) / 2
+
+        # Store the new one
+        k = k + 1
+        tl.store(v + block_start + (k * stride), q)
+        tl.store(z + block_start + (k * stride), s)
+        if k + 1 < length:
+            tl.store(z + block_start + ((k + 1) * stride), 1e9)
+
+    # Last step, we read the envelope to find the min in every location
+    k = 0
+    for q in range(length):
+        while (
+            k + 1 < length
+            and tl.load(
+                z + block_start + ((k + 1) * stride), mask=(k + 1) < length, other=q
+            )
+            < q
+        ):
+            k += 1
+        r = tl.load(v + block_start + (k * stride))
+        d = q - r
+        old_value = tl.load(inputs_ptr + block_start + (r * stride))
+        tl.store(outputs_ptr + block_start + (q * stride), old_value + d * d)
+
+
+def edt_triton(data: torch.Tensor):
+    """
+    Computes the Euclidean Distance Transform (EDT) of a batch of binary images.
+
+    Args:
+        data: A tensor of shape (B, H, W) representing a batch of binary images.
+
+    Returns:
+        A tensor of the same shape as data containing the EDT.
+        It should be equivalent to a batched version of cv2.distanceTransform(input, cv2.DIST_L2, 0)
+    """
+    assert data.dim() == 3
+    assert data.is_cuda
+    B, H, W = data.shape
+    data = data.contiguous()
+
+    # Allocate the "function" tensor. Implicitly the function is 0 if data[i,j]==0 else +infinity
+    output = torch.where(data, 1e18, 0.0)
+    assert output.is_contiguous()
+
+    # Scratch tensors for the parabola stacks
+    parabola_loc = torch.zeros(B, H, W, dtype=torch.uint32, device=data.device)
+    parabola_inter = torch.empty(B, H, W, dtype=torch.float, device=data.device)
+    parabola_inter[:, :, 0] = -1e18
+    parabola_inter[:, :, 1] = 1e18
+
+    # Grid size (number of blocks)
+    grid = (B, H)
+
+    # Launch initialization kernel
+    edt_kernel[grid](
+        output.clone(),
+        output,
+        parabola_loc,
+        parabola_inter,
+        H,
+        W,
+        horizontal=True,
+    )
+
+    # reset the parabola stacks
+    parabola_loc.zero_()
+    parabola_inter[:, :, 0] = -1e18
+    parabola_inter[:, :, 1] = 1e18
+
+    grid = (B, W)
+    edt_kernel[grid](
+        output.clone(),
+        output,
+        parabola_loc,
+        parabola_inter,
+        H,
+        W,
+        horizontal=False,
+    )
+    # don't forget to take sqrt at the end
+    return output.sqrt()

sam3/model/encoder.py

@@ -0,0 +1,594 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+# Based on https://github.com/IDEA-Research/GroundingDINO
+
+from typing import Any, Dict, List, Optional, Tuple
+
+import torch
+from torch import nn, Tensor
+
+from .act_ckpt_utils import activation_ckpt_wrapper
+from .model_misc import get_activation_fn, get_clones, get_valid_ratio
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Transformer encoder layer that performs self-attention followed by cross-attention.
+
+    This layer was previously called TransformerDecoderLayer but was renamed to better
+    reflect its role in the architecture. It processes input sequences through self-attention
+    and then cross-attention with another input (typically image features).
+
+    The layer supports both pre-norm and post-norm configurations, as well as
+    positional encoding at different stages of the attention mechanism.
+    """
+
+    def __init__(
+        self,
+        activation: str,
+        cross_attention: nn.Module,
+        d_model: int,
+        dim_feedforward: int,
+        dropout: float,
+        pos_enc_at_attn: bool,
+        pos_enc_at_cross_attn_keys: bool,
+        pos_enc_at_cross_attn_queries: bool,
+        pre_norm: bool,
+        self_attention: nn.Module,
+    ):
+        """
+        Initialize a transformer encoder layer.
+
+        Args:
+            activation: Activation function to use in the feedforward network
+            cross_attention: Cross-attention module for attending to image features
+            d_model: Model dimension/hidden size
+            dim_feedforward: Dimension of the feedforward network
+            dropout: Dropout probability
+            pos_enc_at_attn: Whether to add positional encodings at self-attention
+            pos_enc_at_cross_attn_keys: Whether to add positional encodings to keys in cross-attention
+            pos_enc_at_cross_attn_queries: Whether to add positional encodings to queries in cross-attention
+            pre_norm: Whether to use pre-norm (True) or post-norm (False) architecture
+            self_attention: Self-attention module
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.dim_feedforward = dim_feedforward
+        self.dropout_value = dropout
+        self.self_attn = self_attention
+        self.cross_attn_image = cross_attention
+
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.activation_str = activation
+        self.activation = get_activation_fn(activation)
+        self.pre_norm = pre_norm
+
+        self.pos_enc_at_attn = pos_enc_at_attn
+        self.pos_enc_at_cross_attn_queries = pos_enc_at_cross_attn_queries
+        self.pos_enc_at_cross_attn_keys = pos_enc_at_cross_attn_keys
+
+        self.layer_idx = None
+
+    def forward_post(
+        self,
+        tgt: Tensor,
+        memory: Tensor,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        **kwargs,
+    ) -> Tensor:
+        """
+        Forward pass for post-norm architecture.
+
+        In post-norm architecture, normalization is applied after attention and feedforward operations.
+
+        Args:
+            tgt: Input tensor to be processed
+            memory: Memory tensor for cross-attention
+            tgt_mask: Mask for self-attention
+            memory_mask: Mask for cross-attention
+            tgt_key_padding_mask: Key padding mask for self-attention
+            memory_key_padding_mask: Key padding mask for cross-attention
+            pos: Positional encoding for memory
+            query_pos: Positional encoding for query
+            **kwargs: Additional keyword arguments
+
+        Returns:
+            Processed tensor
+        """
+        q = k = tgt + query_pos if self.pos_enc_at_attn else tgt
+
+        # Self attention
+        tgt2 = self.self_attn(
+            q, k, value=tgt, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask
+        )[0]
+        tgt = tgt + self.dropout1(tgt2)
+        tgt = self.norm1(tgt)
+
+        # Cross attention to image
+        tgt2 = self.cross_attn_image(
+            query=tgt + query_pos if self.pos_enc_at_cross_attn_queries else tgt,
+            key=memory + pos if self.pos_enc_at_cross_attn_keys else memory,
+            value=memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout2(tgt2)
+        tgt = self.norm2(tgt)
+
+        # FFN
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
+        tgt = tgt + self.dropout3(tgt2)
+        tgt = self.norm3(tgt)
+        return tgt
+
+    def forward_pre(
+        self,
+        tgt: Tensor,
+        memory: Tensor,
+        dac: bool = False,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        # attn_bias: Optional[Tensor] = None,
+        # **kwargs,
+    ) -> Tensor:
+        """
+        Forward pass for pre-norm architecture.
+
+        In pre-norm architecture, normalization is applied before attention and feedforward operations.
+
+        Args:
+            tgt: Input tensor to be processed
+            memory: Memory tensor for cross-attention
+            dac: Whether to use Divide-and-Conquer attention
+            tgt_mask: Mask for self-attention
+            memory_mask: Mask for cross-attention
+            tgt_key_padding_mask: Key padding mask for self-attention
+            memory_key_padding_mask: Key padding mask for cross-attention
+            pos: Positional encoding for memory
+            query_pos: Positional encoding for query
+            attn_bias: Optional attention bias tensor
+            **kwargs: Additional keyword arguments
+
+        Returns:
+            Processed tensor
+        """
+        if dac:
+            # we only apply self attention to the first half of the queries
+            assert tgt.shape[0] % 2 == 0
+            other_tgt = tgt[tgt.shape[0] // 2 :]
+            tgt = tgt[: tgt.shape[0] // 2]
+        tgt2 = self.norm1(tgt)
+        q = k = tgt2 + query_pos if self.pos_enc_at_attn else tgt2
+        tgt2 = self.self_attn(
+            q, k, value=tgt2, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask
+        )[0]
+        tgt = tgt + self.dropout1(tgt2)
+        if dac:
+            # Recombine
+            tgt = torch.cat((tgt, other_tgt), dim=0)
+        tgt2 = self.norm2(tgt)
+        tgt2 = self.cross_attn_image(
+            query=tgt2 + query_pos if self.pos_enc_at_cross_attn_queries else tgt2,
+            key=memory + pos if self.pos_enc_at_cross_attn_keys else memory,
+            value=memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+            # attn_bias=attn_bias,
+        )[0]
+        tgt = tgt + self.dropout2(tgt2)
+        tgt2 = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
+        tgt = tgt + self.dropout3(tgt2)
+        return tgt
+
+    def forward(
+        self,
+        tgt: Tensor,
+        memory: Tensor,
+        dac: bool = False,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        # attn_bias: Optional[Tensor] = None,
+        # **kwds: Any,
+    ) -> torch.Tensor:
+        """
+        Forward pass for the transformer encoder layer.
+
+        Args:
+            tgt: Input tensor to be processed
+            memory: Memory tensor (e.g., image features) for cross-attention
+            dac: Whether to use Divide-and-Conquer attention (only apply self-attention to first half)
+            tgt_mask: Mask for self-attention
+            memory_mask: Mask for cross-attention
+            tgt_key_padding_mask: Key padding mask for self-attention
+            memory_key_padding_mask: Key padding mask for cross-attention
+            pos: Positional encoding for memory
+            query_pos: Positional encoding for query
+            attn_bias: Optional attention bias tensor
+            **kwds: Additional keyword arguments
+
+        Returns:
+            Processed tensor after self-attention, cross-attention, and feedforward network
+        """
+        fwd_fn = self.forward_pre if self.pre_norm else self.forward_post
+        return fwd_fn(
+            tgt,
+            memory,
+            dac=dac,
+            tgt_mask=tgt_mask,
+            memory_mask=memory_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+            pos=pos,
+            query_pos=query_pos,
+            # attn_bias=attn_bias,
+            # **kwds,
+        )
+
+
+class TransformerEncoder(nn.Module):
+    """
+    Transformer encoder that processes multi-level features.
+
+    This encoder takes multi-level features (e.g., from a backbone network) and processes
+    them through a stack of transformer encoder layers. It supports features from multiple
+    levels (e.g., different resolutions) and can apply activation checkpointing for memory
+    efficiency during training.
+
+    Args:
+        layer: The encoder layer to be stacked multiple times
+        num_layers: Number of encoder layers to stack
+        d_model: Model dimension/hidden size
+        num_feature_levels: Number of feature levels to process
+        frozen: Whether to freeze the parameters of this module
+        use_act_checkpoint: Whether to use activation checkpointing during training
+    """
+
+    def __init__(
+        self,
+        layer: nn.Module,
+        num_layers: int,
+        d_model: int,
+        num_feature_levels: int,
+        frozen: bool = False,
+        use_act_checkpoint: bool = False,
+    ):
+        super().__init__()
+        self.layers = get_clones(layer, num_layers)
+        self.num_layers = num_layers
+
+        self.num_feature_levels = num_feature_levels
+        self.level_embed = None
+        if num_feature_levels > 1:
+            self.level_embed = nn.Parameter(torch.Tensor(num_feature_levels, d_model))
+
+        if frozen:
+            for p in self.parameters():
+                p.requires_grad_(False)
+
+        self.use_act_checkpoint = use_act_checkpoint
+
+        # assign layer index to each layer so that some layers can decide what to do
+        # based on which layer index they are (e.g. cross attention to memory bank only
+        # in selected layers)
+        for layer_idx, layer in enumerate(self.layers):
+            layer.layer_idx = layer_idx
+
+    @staticmethod
+    def get_reference_points(spatial_shapes, valid_ratios, device):
+        with torch.no_grad():
+            reference_points_list = []
+            for lvl, (H_, W_) in enumerate(spatial_shapes):
+                ref_y, ref_x = torch.meshgrid(
+                    torch.linspace(
+                        0.5, H_ - 0.5, H_, dtype=torch.float32, device=device
+                    ),
+                    torch.linspace(
+                        0.5, W_ - 0.5, W_, dtype=torch.float32, device=device
+                    ),
+                )
+                ref_y = ref_y.reshape(-1)[None] / (valid_ratios[:, None, lvl, 1] * H_)
+                ref_x = ref_x.reshape(-1)[None] / (valid_ratios[:, None, lvl, 0] * W_)
+                ref = torch.stack((ref_x, ref_y), -1)
+                reference_points_list.append(ref)
+            reference_points = torch.cat(reference_points_list, 1)
+            reference_points = reference_points[:, :, None] * valid_ratios[:, None]
+
+        return reference_points
+
+    def _prepare_multilevel_features(self, srcs, masks, pos_embeds):
+        assert (
+            len(srcs) == self.num_feature_levels
+        ), "mismatch between expected and received # of feature levels"
+
+        src_flatten = []
+        mask_flatten = []
+        lvl_pos_embed_flatten = []
+        spatial_shapes = []
+        has_mask = masks is not None and masks[0] is not None
+        for lvl, (src, mask, pos_embed) in enumerate(zip(srcs, masks, pos_embeds)):
+            bs, c, h, w = src.shape
+            spatial_shape = (h, w)
+            spatial_shapes.append(spatial_shape)
+
+            src = src.flatten(2).transpose(1, 2)  # bs, hw, c
+            if has_mask:
+                mask = mask.flatten(1)
+            pos_embed = pos_embed.flatten(2).transpose(1, 2)  # bs, hw, c
+            if self.level_embed is not None:
+                lvl_pos_embed = pos_embed + self.level_embed[lvl].view(1, 1, -1)
+            else:
+                lvl_pos_embed = pos_embed
+            lvl_pos_embed_flatten.append(lvl_pos_embed)
+            src_flatten.append(src)
+            if has_mask:
+                mask_flatten.append(mask)
+        src_flatten = torch.cat(src_flatten, 1)  # bs, \sum{hxw}, c
+        mask_flatten = torch.cat(mask_flatten, 1) if has_mask else None  # bs, \sum{hxw}
+        lvl_pos_embed_flatten = torch.cat(lvl_pos_embed_flatten, 1)  # bs, \sum{hxw}, c
+        spatial_shapes = torch.tensor(
+            spatial_shapes, dtype=torch.long, device=src_flatten.device
+        )
+        level_start_index = torch.cat(
+            (
+                spatial_shapes.new_zeros((1,)),
+                spatial_shapes.prod(1).cumsum(0)[:-1],
+            )
+        )
+        if has_mask:
+            valid_ratios = torch.stack([get_valid_ratio(m) for m in masks], 1)
+        else:
+            valid_ratios = torch.ones(
+                (src_flatten.shape[0], self.num_feature_levels, 2),
+                device=src_flatten.device,
+            )
+
+        return (
+            src_flatten,
+            mask_flatten,
+            lvl_pos_embed_flatten,
+            level_start_index,
+            valid_ratios,
+            spatial_shapes,
+        )
+
+    def forward(
+        self,
+        src: List[Tensor],
+        src_key_padding_masks: Optional[List[Tensor]] = None,
+        pos: Optional[List[Tensor]] = None,
+        prompt: Optional[Tensor] = None,
+        prompt_key_padding_mask: Optional[Tensor] = None,
+        encoder_extra_kwargs: Optional[Dict] = None,
+    ) -> Tuple[Tensor, Optional[Tensor], Tensor, Tensor, Tensor, Tensor]:
+        """
+        Process multi-level features through the transformer encoder.
+
+        Args:
+            src: List of multi-level features, each with shape (batch_size, channels, height, width)
+            src_key_padding_masks: List of padding masks for each feature level, each with shape (batch_size, height, width)
+            pos: List of positional embeddings for each feature level, each with shape (batch_size, channels, height, width)
+            prompt: Optional text/prompt features to attend to, with shape (seq_len, batch_size, d_model)
+            prompt_key_padding_mask: Optional padding mask for prompt, with shape (batch_size, seq_len)
+            encoder_extra_kwargs: Optional additional arguments to pass to each encoder layer
+
+        Returns:
+            A tuple containing:
+            - output: Processed features with shape (seq_len, batch_size, d_model)
+            - key_padding_masks_flatten: Flattened padding masks
+            - lvl_pos_embed_flatten: Flattened positional embeddings
+            - level_start_index: Starting indices for each feature level
+            - spatial_shapes: Spatial dimensions of each feature level
+            - valid_ratios: Valid ratios for each feature level
+        """
+        assert (
+            len(src) == self.num_feature_levels
+        ), "must be equal to num_feature_levels"
+        if src_key_padding_masks is not None:
+            assert len(src_key_padding_masks) == self.num_feature_levels
+        if pos is not None:
+            assert len(pos) == self.num_feature_levels
+        # Flatten multilevel feats and add level pos embeds
+        (
+            src_flatten,
+            key_padding_masks_flatten,
+            lvl_pos_embed_flatten,
+            level_start_index,
+            valid_ratios,
+            spatial_shapes,
+        ) = self._prepare_multilevel_features(src, src_key_padding_masks, pos)
+
+        reference_points = self.get_reference_points(
+            spatial_shapes, valid_ratios, device=src_flatten.device
+        )
+
+        output = src_flatten
+        for layer in self.layers:
+            layer_kwargs = {}
+
+            assert isinstance(layer, TransformerEncoderLayer)
+            layer_kwargs["memory"] = prompt
+            layer_kwargs["memory_key_padding_mask"] = prompt_key_padding_mask
+            layer_kwargs["query_pos"] = lvl_pos_embed_flatten
+            layer_kwargs["tgt"] = output
+            layer_kwargs["tgt_key_padding_mask"] = key_padding_masks_flatten
+
+            if self.training:
+                assert self.use_act_checkpoint, "activation ckpt not enabled in encoder"
+            if encoder_extra_kwargs is not None:
+                layer_kwargs.update(encoder_extra_kwargs)
+            output = activation_ckpt_wrapper(layer)(
+                **layer_kwargs,
+                act_ckpt_enable=self.training and self.use_act_checkpoint,
+            )
+        # return as seq first
+        return (
+            output.transpose(0, 1),
+            (
+                key_padding_masks_flatten.transpose(0, 1)
+                if key_padding_masks_flatten is not None
+                else None
+            ),
+            lvl_pos_embed_flatten.transpose(0, 1),
+            level_start_index,
+            spatial_shapes,
+            valid_ratios,
+        )
+
+
+class TransformerEncoderFusion(TransformerEncoder):
+    """
+    Transformer encoder that fuses text and image features.
+
+    This encoder extends TransformerEncoder to handle both text and image features,
+    with the ability to add pooled text features to image features for better
+    cross-modal fusion. It supports torch.compile for performance optimization.
+
+    Args:
+        layer: The encoder layer to be stacked multiple times
+        num_layers: Number of encoder layers to stack
+        d_model: Model dimension/hidden size
+        num_feature_levels: Number of feature levels to process
+        add_pooled_text_to_img_feat: Whether to add pooled text features to image features
+        pool_text_with_mask: Whether to use the mask when pooling text features
+        compile_mode: Mode for torch.compile, or None to disable compilation
+        **kwargs: Additional arguments to pass to the parent class
+    """
+
+    def __init__(
+        self,
+        layer: nn.Module,
+        num_layers: int,
+        d_model: int,
+        num_feature_levels: int,
+        add_pooled_text_to_img_feat: bool = True,
+        pool_text_with_mask: bool = False,
+        compile_mode: Optional[str] = None,
+        **kwargs,
+    ):
+        super().__init__(
+            layer,
+            num_layers,
+            d_model,
+            num_feature_levels,
+            **kwargs,
+        )
+        self.add_pooled_text_to_img_feat = add_pooled_text_to_img_feat
+        if self.add_pooled_text_to_img_feat:
+            self.text_pooling_proj = nn.Linear(d_model, d_model)
+        self.pool_text_with_mask = pool_text_with_mask
+        if compile_mode is not None:
+            self.forward = torch.compile(
+                self.forward, mode=compile_mode, fullgraph=True
+            )
+
+    @staticmethod
+    def get_reference_points(spatial_shapes, valid_ratios, device):
+        # Not needed here
+        return None
+
+    def forward(
+        self,
+        src: List[Tensor],
+        prompt: Tensor,
+        src_key_padding_mask: Optional[List[Tensor]] = None,
+        src_pos: Optional[List[Tensor]] = None,
+        prompt_key_padding_mask: Optional[Tensor] = None,
+        prompt_pos: Optional[Tensor] = None,
+        feat_sizes: Optional[List[int]] = None,
+        encoder_extra_kwargs: Optional[Dict] = None,
+    ):
+        # Restore spatial shapes of vision
+        bs = src[0].shape[1]  # seq first
+        if feat_sizes is not None:
+            assert len(feat_sizes) == len(src)
+            if src_key_padding_mask is None:
+                src_key_padding_mask = [None] * len(src)
+            for i, (h, w) in enumerate(feat_sizes):
+                src[i] = src[i].reshape(h, w, bs, -1).permute(2, 3, 0, 1)
+                src_pos[i] = src_pos[i].reshape(h, w, bs, -1).permute(2, 3, 0, 1)
+                src_key_padding_mask[i] = (
+                    src_key_padding_mask[i].reshape(h, w, bs).permute(2, 0, 1)
+                    if src_key_padding_mask[i] is not None
+                    else None
+                )
+        else:
+            assert all(
+                x.dim == 4 for x in src
+            ), "expected list of (bs, c, h, w) tensors"
+
+        if self.add_pooled_text_to_img_feat:
+            # Fusion: Add mean pooled text to image features
+            pooled_text = pool_text_feat(
+                prompt, prompt_key_padding_mask, self.pool_text_with_mask
+            )
+            pooled_text = self.text_pooling_proj(pooled_text)[
+                ..., None, None
+            ]  # prompt is seq first
+            src = [x.add_(pooled_text) for x in src]
+
+        (
+            out,
+            key_padding_masks_flatten,
+            lvl_pos_embed_flatten,
+            level_start_index,
+            spatial_shapes,
+            valid_ratios,
+        ) = super().forward(
+            src,
+            src_key_padding_masks=src_key_padding_mask,
+            pos=src_pos,
+            prompt=prompt.transpose(0, 1),
+            prompt_key_padding_mask=prompt_key_padding_mask,
+            encoder_extra_kwargs=encoder_extra_kwargs,
+        )
+
+        return {
+            "memory": out,
+            "padding_mask": key_padding_masks_flatten,
+            "pos_embed": lvl_pos_embed_flatten,
+            "memory_text": prompt,
+            "level_start_index": level_start_index,
+            "spatial_shapes": spatial_shapes,
+            "valid_ratios": valid_ratios,
+        }
+
+
+def pool_text_feat(prompt, prompt_mask, pool_with_mask):
+    # prompt has shape (seq, bs, dim)
+    if not pool_with_mask:
+        return prompt.mean(dim=0)
+
+    # prompt_mask has shape (bs, seq), where False is valid and True is padding
+    assert prompt_mask.dim() == 2
+    # is_valid has shape (seq, bs, 1), where 1 is valid and 0 is padding
+    is_valid = (~prompt_mask).float().permute(1, 0)[..., None]
+    # num_valid has shape (bs, 1)
+    num_valid = torch.clamp(torch.sum(is_valid, dim=0), min=1.0)
+
+    # mean pool over all the valid tokens
+    pooled_text = (prompt * is_valid).sum(dim=0) / num_valid
+    return pooled_text

sam3/model/geometry_encoders.py

@@ -0,0 +1,850 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+import torchvision
+from typing_extensions import override
+
+from .act_ckpt_utils import activation_ckpt_wrapper
+from .box_ops import box_cxcywh_to_xyxy
+
+from .model_misc import get_clones
+
+
+def is_right_padded(mask):
+    """Given a padding mask (following pytorch convention, 1s for padded values),
+    returns whether the padding is on the right or not."""
+    return (mask.long() == torch.sort(mask.long(), dim=-1)[0]).all()
+
+
+def concat_padded_sequences(seq1, mask1, seq2, mask2, return_index: bool = False):
+    """
+    Concatenates two right-padded sequences, such that the resulting sequence
+    is contiguous and also right-padded.
+
+    Following pytorch's convention, tensors are sequence first, and the mask are
+    batch first, with 1s for padded values.
+
+    :param seq1: A tensor of shape (seq1_length, batch_size, hidden_size).
+    :param mask1: A tensor of shape (batch_size, seq1_length).
+    :param seq2: A tensor of shape (seq2_length, batch_size,  hidden_size).
+    :param mask2: A tensor of shape (batch_size, seq2_length).
+    :param return_index: If True, also returns the index of the ids of the element of seq2
+        in the concatenated sequence. This can be used to retrieve the elements of seq2
+    :return: A tuple (concatenated_sequence, concatenated_mask) if return_index is False,
+        otherwise (concatenated_sequence, concatenated_mask, index).
+    """
+    seq1_length, batch_size, hidden_size = seq1.shape
+    seq2_length, batch_size, hidden_size = seq2.shape
+
+    assert batch_size == seq1.size(1) == seq2.size(1) == mask1.size(0) == mask2.size(0)
+    assert hidden_size == seq1.size(2) == seq2.size(2)
+    assert seq1_length == mask1.size(1)
+    assert seq2_length == mask2.size(1)
+
+    torch._assert_async(is_right_padded(mask1))
+    torch._assert_async(is_right_padded(mask2))
+
+    actual_seq1_lengths = (~mask1).sum(dim=-1)
+    actual_seq2_lengths = (~mask2).sum(dim=-1)
+
+    final_lengths = actual_seq1_lengths + actual_seq2_lengths
+    max_length = seq1_length + seq2_length
+    concatenated_mask = (
+        torch.arange(max_length, device=seq2.device)[None].repeat(batch_size, 1)
+        >= final_lengths[:, None]
+    )
+
+    # (max_len, batch_size, hidden_size)
+    concatenated_sequence = torch.zeros(
+        (max_length, batch_size, hidden_size), device=seq2.device, dtype=seq2.dtype
+    )
+    concatenated_sequence[:seq1_length, :, :] = seq1
+
+    # At this point, the element of seq1 are in the right place
+    # We just need to shift the elements of seq2
+
+    index = torch.arange(seq2_length, device=seq2.device)[:, None].repeat(1, batch_size)
+    index = index + actual_seq1_lengths[None]
+
+    concatenated_sequence = concatenated_sequence.scatter(
+        0, index[:, :, None].expand(-1, -1, hidden_size), seq2
+    )
+
+    if return_index:
+        return concatenated_sequence, concatenated_mask, index
+
+    return concatenated_sequence, concatenated_mask
+
+
+class Prompt:
+    """Utility class to manipulate geometric prompts.
+
+    We expect the sequences in pytorch convention, that is sequence first, batch second
+    The dimensions are expected as follows:
+    box_embeddings shape: N_boxes x B x C_box
+    box_mask shape: B x N_boxes. Can be None if nothing is masked out
+    point_embeddings shape: N_points x B x C_point
+    point_mask shape: B x N_points. Can be None if nothing is masked out
+    mask_embeddings shape: N_masks x B x 1 x H_mask x W_mask
+    mask_mask shape: B x N_masks. Can be None if nothing is masked out
+
+    We also store positive/negative labels. These tensors are also stored batch-first
+    If they are None, we'll assume positive labels everywhere
+    box_labels: long tensor of shape N_boxes x B
+    point_labels: long tensor of shape N_points x B
+    mask_labels: long tensor of shape N_masks x B
+    """
+
+    def __init__(
+        self,
+        box_embeddings=None,
+        box_mask=None,
+        point_embeddings=None,
+        point_mask=None,
+        box_labels=None,
+        point_labels=None,
+        mask_embeddings=None,
+        mask_mask=None,  # Attention mask for mask prompt
+        mask_labels=None,
+    ):
+        # Check for null prompt
+        if (
+            box_embeddings is None
+            and point_embeddings is None
+            and mask_embeddings is None
+        ):
+            self.box_embeddings = None
+            self.box_labels = None
+            self.box_mask = None
+            self.point_embeddings = None
+            self.point_labels = None
+            self.point_mask = None
+            self.mask_embeddings = None
+            self.mask_mask = None
+            # Masks are assumed positive only for now.
+            self.mask_labels = None
+            return
+        # Get sequence lengths and device
+        box_seq_len, point_seq_len, mask_seq_len, bs, device = (
+            self._init_seq_len_and_device(
+                box_embeddings, point_embeddings, mask_embeddings
+            )
+        )
+
+        # Initialize embeds, labels, attention masks.
+        box_embeddings, box_labels, box_mask = self._init_box(
+            box_embeddings, box_labels, box_mask, box_seq_len, bs, device
+        )
+        point_embeddings, point_labels, point_mask = self._init_point(
+            point_embeddings, point_labels, point_mask, point_seq_len, bs, device
+        )
+        mask_embeddings, mask_labels, mask_mask = self._init_mask(
+            mask_embeddings, mask_labels, mask_mask, mask_seq_len, bs, device
+        )
+
+        # Dimension checks
+        assert (
+            box_embeddings is not None
+            and list(box_embeddings.shape[:2])
+            == [
+                box_seq_len,
+                bs,
+            ]
+        ), f"Wrong dimension for box embeddings. Expected [{box_seq_len}, {bs}, *] got {box_embeddings.shape}"
+        assert (
+            box_mask is not None
+            and list(box_mask.shape)
+            == [
+                bs,
+                box_seq_len,
+            ]
+        ), f"Wrong dimension for box mask. Expected [{bs}, {box_seq_len}] got {box_mask.shape}"
+        assert (
+            point_embeddings is not None
+            and list(point_embeddings.shape[:2])
+            == [
+                point_seq_len,
+                bs,
+            ]
+        ), f"Wrong dimension for point embeddings. Expected [{point_seq_len}, {bs}, *] got {point_embeddings.shape}"
+        assert (
+            point_mask is not None
+            and list(point_mask.shape)
+            == [
+                bs,
+                point_seq_len,
+            ]
+        ), f"Wrong dimension for point mask. Expected [{bs}, {point_seq_len}] got {point_mask.shape}"
+        assert (
+            box_labels is not None
+            and list(box_labels.shape)
+            == [
+                box_seq_len,
+                bs,
+            ]
+        ), f"Wrong dimension for box labels. Expected [{box_seq_len}, {bs}] got {box_labels.shape}"
+        assert (
+            point_labels is not None
+            and list(point_labels.shape)
+            == [
+                point_seq_len,
+                bs,
+            ]
+        ), f"Wrong dimension for point labels. Expected [{point_seq_len}, {bs}] got {point_labels.shape}"
+        assert (
+            # Allowed to be None, we leave it to the encoder to check for validity before encoding.
+            mask_embeddings is None
+            or list(mask_embeddings.shape[:2])
+            == [
+                mask_seq_len,
+                bs,
+            ]
+        ), f"Wrong dimension for mask embeddings. Expected [{mask_seq_len}, {bs}, *] got {mask_embeddings.shape}"
+        assert (
+            mask_mask is None
+            or list(mask_mask.shape)
+            == [
+                bs,
+                mask_seq_len,
+            ]
+        ), f"Wrong dimension for mask attn. mask. Expected [{bs}, {mask_seq_len}] got {mask_mask.shape}"
+
+        # Device checks
+        assert (
+            box_embeddings is not None and box_embeddings.device == device
+        ), f"Expected box embeddings to be on device {device}, got {box_embeddings.device}"
+        assert (
+            box_mask is not None and box_mask.device == device
+        ), f"Expected box mask to be on device {device}, got {box_mask.device}"
+        assert (
+            box_labels is not None and box_labels.device == device
+        ), f"Expected box labels to be on device {device}, got {box_labels.device}"
+        assert (
+            point_embeddings is not None and point_embeddings.device == device
+        ), f"Expected point embeddings to be on device {device}, got {point_embeddings.device}"
+        assert (
+            point_mask is not None and point_mask.device == device
+        ), f"Expected point mask to be on device {device}, got {point_mask.device}"
+        assert (
+            point_labels is not None and point_labels.device == device
+        ), f"Expected point labels to be on device {device}, got {point_labels.device}"
+        assert (
+            mask_embeddings is None or mask_embeddings.device == device
+        ), f"Expected mask embeddings to be on device {device}, got {mask_embeddings.device}"
+        assert (
+            mask_mask is None or mask_mask.device == device
+        ), f"Expected mask attn. mask to be on device {device}, got {mask_mask.device}"
+
+        self.box_embeddings = box_embeddings
+        self.point_embeddings = point_embeddings
+        self.box_mask = box_mask
+        self.point_mask = point_mask
+        self.box_labels = box_labels
+        self.point_labels = point_labels
+        self.mask_embeddings = mask_embeddings
+        self.mask_labels = mask_labels
+        self.mask_mask = mask_mask
+
+    def _init_seq_len_and_device(
+        self, box_embeddings, point_embeddings, mask_embeddings
+    ):
+        box_seq_len = point_seq_len = mask_seq_len = 0
+        bs = None
+        device = None
+        if box_embeddings is not None:
+            bs = box_embeddings.shape[1]
+            box_seq_len = box_embeddings.shape[0]
+            device = box_embeddings.device
+
+        if point_embeddings is not None:
+            point_seq_len = point_embeddings.shape[0]
+            if bs is not None:
+                assert (
+                    bs == point_embeddings.shape[1]
+                ), f"Batch size mismatch between box and point embeddings. Got {bs} and {point_embeddings.shape[1]}."
+            else:
+                bs = point_embeddings.shape[1]
+            if device is not None:
+                assert (
+                    device == point_embeddings.device
+                ), "Device mismatch between box and point embeddings"
+            else:
+                device = point_embeddings.device
+
+        if mask_embeddings is not None:
+            mask_seq_len = mask_embeddings.shape[0]
+            if bs is not None:
+                assert (
+                    bs == mask_embeddings.shape[1]
+                ), f"Batch size mismatch between box/point and mask embedding. Got {bs} and {mask_embeddings.shape[1]}"
+            else:
+                bs = mask_embeddings.shape[1]
+            if device is not None:
+                assert (
+                    device == mask_embeddings.device
+                ), "Device mismatch between box/point and mask embeddings."
+            else:
+                device = mask_embeddings.device
+
+        return box_seq_len, point_seq_len, mask_seq_len, bs, device
+
+    def _init_box(self, box_embeddings, box_labels, box_mask, box_seq_len, bs, device):
+        if box_embeddings is None:
+            box_embeddings = torch.zeros(box_seq_len, bs, 4, device=device)
+        if box_labels is None:
+            box_labels = torch.ones(box_seq_len, bs, device=device, dtype=torch.long)
+        if box_mask is None:
+            box_mask = torch.zeros(bs, box_seq_len, device=device, dtype=torch.bool)
+        return box_embeddings, box_labels, box_mask
+
+    def _init_point(
+        self, point_embeddings, point_labels, point_mask, point_seq_len, bs, device
+    ):
+        """
+        Identical to _init_box. Except that C=2 for points (vs. 4 for boxes).
+        """
+        if point_embeddings is None:
+            point_embeddings = torch.zeros(point_seq_len, bs, 2, device=device)
+        if point_labels is None:
+            point_labels = torch.ones(
+                point_seq_len, bs, device=device, dtype=torch.long
+            )
+        if point_mask is None:
+            point_mask = torch.zeros(bs, point_seq_len, device=device, dtype=torch.bool)
+        return point_embeddings, point_labels, point_mask
+
+    def _init_mask(
+        self, mask_embeddings, mask_labels, mask_mask, mask_seq_len, bs, device
+    ):
+        # NOTE: Mask embeddings can be of arbitrary resolution, so we don't initialize it here.
+        # In case we append new mask, we check that its resolution matches exisiting ones (if any).
+        # In case mask_embeddings is None, we should never encode it.
+        if mask_labels is None:
+            mask_labels = torch.ones(mask_seq_len, bs, device=device, dtype=torch.long)
+        if mask_mask is None:
+            mask_mask = torch.zeros(bs, mask_seq_len, device=device, dtype=torch.bool)
+        return mask_embeddings, mask_labels, mask_mask
+
+    def append_boxes(self, boxes, labels, mask=None):
+        if self.box_embeddings is None:
+            self.box_embeddings = boxes
+            self.box_labels = labels
+            self.box_mask = mask
+            return
+
+        bs = self.box_embeddings.shape[1]
+        assert boxes.shape[1] == labels.shape[1] == bs
+        assert list(boxes.shape[:2]) == list(labels.shape[:2])
+        if mask is None:
+            mask = torch.zeros(
+                bs, boxes.shape[0], dtype=torch.bool, device=boxes.device
+            )
+
+        self.box_labels, _ = concat_padded_sequences(
+            self.box_labels.unsqueeze(-1), self.box_mask, labels.unsqueeze(-1), mask
+        )
+        self.box_labels = self.box_labels.squeeze(-1)
+        self.box_embeddings, self.box_mask = concat_padded_sequences(
+            self.box_embeddings, self.box_mask, boxes, mask
+        )
+
+    def append_points(self, points, labels, mask=None):
+        if self.point_embeddings is None:
+            self.point_embeddings = points
+            self.point_labels = labels
+            self.point_mask = mask
+            return
+
+        bs = self.point_embeddings.shape[1]
+        assert points.shape[1] == labels.shape[1] == bs
+        assert list(points.shape[:2]) == list(labels.shape[:2])
+        if mask is None:
+            mask = torch.zeros(
+                bs, points.shape[0], dtype=torch.bool, device=points.device
+            )
+
+        self.point_labels, _ = concat_padded_sequences(
+            self.point_labels.unsqueeze(-1), self.point_mask, labels.unsqueeze(-1), mask
+        )
+        self.point_labels = self.point_labels.squeeze(-1)
+        self.point_embeddings, self.point_mask = concat_padded_sequences(
+            self.point_embeddings, self.point_mask, points, mask
+        )
+
+    def append_masks(self, masks, labels=None, attn_mask=None):
+        if labels is not None:
+            assert list(masks.shape[:2]) == list(labels.shape[:2])
+        if self.mask_embeddings is None:
+            self.mask_embeddings = masks
+            mask_seq_len, bs = masks.shape[:2]
+            if labels is None:
+                self.mask_labels = torch.ones(
+                    mask_seq_len, bs, device=masks.device, dtype=torch.long
+                )
+            else:
+                self.mask_labels = labels
+            if attn_mask is None:
+                self.mask_mask = torch.zeros(
+                    bs, mask_seq_len, device=masks.device, dtype=torch.bool
+                )
+            else:
+                self.mask_mask = attn_mask
+        else:
+            raise NotImplementedError("Only one mask per prompt is supported.")
+
+    def clone(self):
+        return Prompt(
+            box_embeddings=(
+                None if self.box_embeddings is None else self.box_embeddings.clone()
+            ),
+            box_mask=None if self.box_mask is None else self.box_mask.clone(),
+            point_embeddings=(
+                None if self.point_embeddings is None else self.point_embeddings.clone()
+            ),
+            point_mask=None if self.point_mask is None else self.point_mask.clone(),
+            box_labels=None if self.box_labels is None else self.box_labels.clone(),
+            point_labels=(
+                None if self.point_labels is None else self.point_labels.clone()
+            ),
+        )
+
+
+class MaskEncoder(nn.Module):
+    """
+    Base class for mask encoders.
+    """
+
+    def __init__(
+        self,
+        mask_downsampler: nn.Module,
+        position_encoding: nn.Module,
+    ):
+        super().__init__()
+        self.mask_downsampler = mask_downsampler
+        self.position_encoding = position_encoding
+
+    def forward(self, masks, *args, **kwargs) -> Tuple[torch.Tensor, torch.Tensor]:
+        masks = self.mask_downsampler(masks)
+        masks_pos = self.position_encoding(masks).to(masks.dtype)
+
+        return masks, masks_pos
+
+
+class FusedMaskEncoder(MaskEncoder):
+    """
+    Identical to memory.SimpleMaskEncoder but follows the interface of geometry_encoders.MaskEncoder.
+    We also remove the `skip_mask_sigmoid` option (to be handled outside the MaskEncoder).
+    Fuses backbone image features with mask features.
+    """
+
+    def __init__(
+        self,
+        mask_downsampler: nn.Module,
+        position_encoding: nn.Module,
+        fuser: nn.Module,
+        in_dim: int = 256,
+        out_dim: int = 256,
+    ):
+        super().__init__(mask_downsampler, position_encoding)
+        self.fuser = fuser
+        self.out_proj = nn.Identity()
+        if out_dim != in_dim:
+            self.out_proj = nn.Conv2d(in_dim, out_dim, kernel_size=1)
+        self.pix_feat_proj = nn.Conv2d(in_dim, in_dim, kernel_size=1)
+
+    @override
+    def forward(
+        self,
+        masks: torch.Tensor,
+        pix_feat: torch.Tensor,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        masks = self.mask_downsampler(masks)
+
+        ## Fuse pix_feats and downsampled masks
+        # in case the visual features are on CPU, cast them to CUDA
+        pix_feat = pix_feat.to(masks.device)
+
+        x = self.pix_feat_proj(pix_feat)
+        x = x + masks
+        x = self.fuser(x)
+        x = self.out_proj(x)
+
+        pos = self.position_encoding(x).to(x.dtype)
+
+        return x, pos
+
+
+class SequenceGeometryEncoder(nn.Module):
+    """
+    This a fully fledged encoder for geometric prompts.
+    It assumes boxes are passed in the "normalized CxCyWH" format, and points in normalized xy
+    This allows flexibility in how to encode the features (eg do pooling)
+
+    Points and boxes can be encoded with any of the three possibilities:
+     - direct projection: we just compute a linear from coordinate space to d_model
+     - pooling: pool features from the backbone in the requested location.
+                For boxes, it's a roi align
+                For points it's a grid sample
+     - pos encoder: Take the position encoding of the point or box center
+
+    These three options are mutually compatible. If several are selected, we'll take a simple addition
+
+    As an alternative, we offer the possibility to encode points only.
+    In that case, the boxes are converted to two points for the top left and bottom right corners (with appropriate labels)
+
+    On top of these encodings, we offer the possibility to further encode the prompt sequence with a transformer.
+    """
+
+    def __init__(
+        self,
+        encode_boxes_as_points: bool,
+        points_direct_project: bool,
+        points_pool: bool,
+        points_pos_enc: bool,
+        boxes_direct_project: bool,
+        boxes_pool: bool,
+        boxes_pos_enc: bool,
+        d_model: int,
+        pos_enc,
+        num_layers: int,
+        layer: nn.Module,
+        roi_size: int = 7,  # for boxes pool
+        add_cls: bool = True,
+        add_post_encode_proj: bool = True,
+        mask_encoder: MaskEncoder = None,
+        add_mask_label: bool = False,
+        use_act_ckpt: bool = False,
+    ):
+        super().__init__()
+
+        self.d_model = d_model
+        self.pos_enc = pos_enc
+        self.encode_boxes_as_points = encode_boxes_as_points
+        self.roi_size = roi_size
+        # There usually are two labels: positive and negatives.
+        # If we encode boxes as points, we have 3 types of points: regular, top left, bottom right
+        # These 3 types can be positives or negatives, hence 2*3 = 6 labels
+        num_labels = 6 if self.encode_boxes_as_points else 2
+        self.label_embed = torch.nn.Embedding(num_labels, self.d_model)
+
+        # This is a cls token, can be used for pooling if need be.
+        # It also ensures that the encoded sequences are always non-empty
+        self.cls_embed = None
+        if add_cls:
+            self.cls_embed = torch.nn.Embedding(1, self.d_model)
+
+        assert (
+            points_direct_project or points_pos_enc or points_pool
+        ), "Error: need at least one way to encode points"
+        assert (
+            encode_boxes_as_points
+            or boxes_direct_project
+            or boxes_pos_enc
+            or boxes_pool
+        ), "Error: need at least one way to encode boxes"
+
+        self.points_direct_project = None
+        if points_direct_project:
+            self.points_direct_project = nn.Linear(2, self.d_model)
+        self.points_pool_project = None
+        if points_pool:
+            self.points_pool_project = nn.Linear(self.d_model, self.d_model)
+        self.points_pos_enc_project = None
+        if points_pos_enc:
+            self.points_pos_enc_project = nn.Linear(self.d_model, self.d_model)
+
+        self.boxes_direct_project = None
+        self.boxes_pool_project = None
+        self.boxes_pos_enc_project = None
+        if not encode_boxes_as_points:
+            if boxes_direct_project:
+                self.boxes_direct_project = nn.Linear(4, self.d_model)
+            if boxes_pool:
+                self.boxes_pool_project = nn.Conv2d(
+                    self.d_model, self.d_model, self.roi_size
+                )
+            if boxes_pos_enc:
+                self.boxes_pos_enc_project = nn.Linear(self.d_model + 2, self.d_model)
+
+        self.final_proj = None
+        if add_post_encode_proj:
+            self.final_proj = nn.Linear(self.d_model, self.d_model)
+            self.norm = nn.LayerNorm(self.d_model)
+
+        self.img_pre_norm = nn.Identity()
+        if self.points_pool_project is not None or self.boxes_pool_project is not None:
+            self.img_pre_norm = nn.LayerNorm(self.d_model)
+
+        self.encode = None
+        if num_layers > 0:
+            assert (
+                add_cls
+            ), "It's currently highly recommended to add a CLS when using a transformer"
+            self.encode = get_clones(layer, num_layers)
+            self.encode_norm = nn.LayerNorm(self.d_model)
+
+        if mask_encoder is not None:
+            assert isinstance(
+                mask_encoder, MaskEncoder
+            ), f"Expected mask_encoder of type MaskEncoder. Got {type(mask_encoder)}."
+            if add_mask_label:
+                self.mask_label_embed = torch.nn.Embedding(2, self.d_model)
+        self.add_mask_label = add_mask_label
+        self.mask_encoder = mask_encoder
+        self.use_act_ckpt = use_act_ckpt
+
+    def _encode_points(self, points, points_mask, points_labels, img_feats):
+        points_embed = None
+        n_points, bs = points.shape[:2]
+
+        if self.points_direct_project is not None:
+            proj = self.points_direct_project(points)
+            assert points_embed is None
+            points_embed = proj
+
+        if self.points_pool_project is not None:
+            # points are [Num_points, bs, 2], normalized in [0, 1]
+            # the grid needs to be [Bs, H_out, W_out, 2] normalized in [-1,1]
+            # Will take H_out = num_points, w_out = 1
+            grid = points.transpose(0, 1).unsqueeze(2)
+            # re normalize to [-1, 1]
+            grid = (grid * 2) - 1
+            sampled = torch.nn.functional.grid_sample(
+                img_feats, grid, align_corners=False
+            )
+            assert list(sampled.shape) == [bs, self.d_model, n_points, 1]
+            sampled = sampled.squeeze(-1).permute(2, 0, 1)
+            proj = self.points_pool_project(sampled)
+            if points_embed is None:
+                points_embed = proj
+            else:
+                points_embed = points_embed + proj
+
+        if self.points_pos_enc_project is not None:
+            x, y = points.unbind(-1)
+            enc_x, enc_y = self.pos_enc._encode_xy(x.flatten(), y.flatten())
+            enc_x = enc_x.view(n_points, bs, enc_x.shape[-1])
+            enc_y = enc_y.view(n_points, bs, enc_y.shape[-1])
+            enc = torch.cat([enc_x, enc_y], -1)
+
+            proj = self.points_pos_enc_project(enc)
+            if points_embed is None:
+                points_embed = proj
+            else:
+                points_embed = points_embed + proj
+
+        type_embed = self.label_embed(points_labels.long())
+        return type_embed + points_embed, points_mask
+
+    def _encode_boxes(self, boxes, boxes_mask, boxes_labels, img_feats):
+        boxes_embed = None
+        n_boxes, bs = boxes.shape[:2]
+
+        if self.boxes_direct_project is not None:
+            proj = self.boxes_direct_project(boxes)
+            assert boxes_embed is None
+            boxes_embed = proj
+
+        if self.boxes_pool_project is not None:
+            H, W = img_feats.shape[-2:]
+
+            # boxes are [Num_boxes, bs, 4], normalized in [0, 1]
+            # We need to denormalize, and convert to [x, y, x, y]
+            boxes_xyxy = box_cxcywh_to_xyxy(boxes)
+            scale = torch.tensor([W, H, W, H], dtype=boxes_xyxy.dtype)
+            scale = scale.pin_memory().to(device=boxes_xyxy.device, non_blocking=True)
+            scale = scale.view(1, 1, 4)
+            boxes_xyxy = boxes_xyxy * scale
+            sampled = torchvision.ops.roi_align(
+                img_feats, boxes_xyxy.float().transpose(0, 1).unbind(0), self.roi_size
+            )
+            assert list(sampled.shape) == [
+                bs * n_boxes,
+                self.d_model,
+                self.roi_size,
+                self.roi_size,
+            ]
+            proj = self.boxes_pool_project(sampled)
+            proj = proj.view(bs, n_boxes, self.d_model).transpose(0, 1)
+            if boxes_embed is None:
+                boxes_embed = proj
+            else:
+                boxes_embed = boxes_embed + proj
+
+        if self.boxes_pos_enc_project is not None:
+            cx, cy, w, h = boxes.unbind(-1)
+            enc = self.pos_enc.encode_boxes(
+                cx.flatten(), cy.flatten(), w.flatten(), h.flatten()
+            )
+            enc = enc.view(boxes.shape[0], boxes.shape[1], enc.shape[-1])
+
+            proj = self.boxes_pos_enc_project(enc)
+            if boxes_embed is None:
+                boxes_embed = proj
+            else:
+                boxes_embed = boxes_embed + proj
+
+        type_embed = self.label_embed(boxes_labels.long())
+        return type_embed + boxes_embed, boxes_mask
+
+    def _encode_masks(
+        self,
+        masks: torch.Tensor,
+        attn_mask: torch.Tensor,
+        mask_labels: torch.Tensor,
+        img_feats: torch.Tensor = None,
+    ):
+        n_masks, bs = masks.shape[:2]
+        assert (
+            n_masks == 1
+        ), "We assume one mask per prompt for now. Code should still be functional if this assertion is removed."
+        assert (
+            list(attn_mask.shape)
+            == [
+                bs,
+                n_masks,
+            ]
+        ), f"Expected attn_mask to be of shape {bs}x{n_masks}. Got {list(attn_mask.shape)}."
+        masks, pos = self.mask_encoder(
+            masks=masks.flatten(0, 1).float(),
+            pix_feat=img_feats,
+        )
+        H, W = masks.shape[-2:]
+        n_tokens_per_mask = H * W
+        # NOTE: We directly add pos enc here as we usually don't keep track of pos encoding for the concatenated prompt (text, other geometric prompts). Might need to do some refactoring for more flexibility.
+        masks = masks + pos
+        masks = masks.view(n_masks, bs, *masks.shape[1:]).flatten(
+            -2
+        )  # n_masks x bs x C x H*W
+        masks = masks.permute(0, 3, 1, 2).flatten(0, 1)  # n_masks * H*W x bs x C
+        attn_mask = attn_mask.repeat_interleave(n_tokens_per_mask, dim=1)
+        if self.add_mask_label:
+            masks = masks + self.mask_label_embed(mask_labels.long())
+        return masks, attn_mask
+
+    def forward(self, geo_prompt: Prompt, img_feats, img_sizes, img_pos_embeds=None):
+        points = geo_prompt.point_embeddings
+        points_mask = geo_prompt.point_mask
+        points_labels = geo_prompt.point_labels
+        boxes = geo_prompt.box_embeddings
+        boxes_mask = geo_prompt.box_mask
+        boxes_labels = geo_prompt.box_labels
+        masks = geo_prompt.mask_embeddings
+        masks_mask = geo_prompt.mask_mask
+        masks_labels = geo_prompt.mask_labels
+        seq_first_img_feats = img_feats[-1]  # [H*W, B, C]
+        seq_first_img_pos_embeds = (
+            img_pos_embeds[-1]
+            if img_pos_embeds is not None
+            else torch.zeros_like(seq_first_img_feats)
+        )
+
+        if self.points_pool_project or self.boxes_pool_project:
+            assert len(img_feats) == len(img_sizes)
+            cur_img_feat = img_feats[-1]
+            cur_img_feat = self.img_pre_norm(cur_img_feat)
+            H, W = img_sizes[-1]
+            assert cur_img_feat.shape[0] == H * W
+            N, C = cur_img_feat.shape[-2:]
+            # Put back in NxCxHxW
+            cur_img_feat = cur_img_feat.permute(1, 2, 0)
+            cur_img_feat = cur_img_feat.view(N, C, H, W)
+            img_feats = cur_img_feat
+
+        if self.encode_boxes_as_points:
+            assert boxes is not None
+            assert geo_prompt.box_mask is not None
+            assert geo_prompt.box_labels is not None
+            assert boxes.shape[-1] == 4
+
+            boxes_xyxy = box_cxcywh_to_xyxy(boxes)
+            top_left, bottom_right = boxes_xyxy.split(split_size=2, dim=-1)
+
+            labels_tl = geo_prompt.box_labels + 2
+            labels_br = geo_prompt.box_labels + 4
+
+            # Append to the existing points
+            points, _ = concat_padded_sequences(
+                points, points_mask, top_left, boxes_mask
+            )
+            points_labels, points_mask = concat_padded_sequences(
+                points_labels.unsqueeze(-1),
+                points_mask,
+                labels_tl.unsqueeze(-1),
+                boxes_mask,
+            )
+            points_labels = points_labels.squeeze(-1)
+
+            points, _ = concat_padded_sequences(
+                points, points_mask, bottom_right, boxes_mask
+            )
+            points_labels, points_mask = concat_padded_sequences(
+                points_labels.unsqueeze(-1),
+                points_mask,
+                labels_br.unsqueeze(-1),
+                boxes_mask,
+            )
+            points_labels = points_labels.squeeze(-1)
+
+        final_embeds, final_mask = self._encode_points(
+            points=points,
+            points_mask=points_mask,
+            points_labels=points_labels,
+            img_feats=img_feats,
+        )
+
+        if not self.encode_boxes_as_points:
+            boxes_embeds, boxes_mask = self._encode_boxes(
+                boxes=boxes,
+                boxes_mask=boxes_mask,
+                boxes_labels=boxes_labels,
+                img_feats=img_feats,
+            )
+
+            final_embeds, final_mask = concat_padded_sequences(
+                final_embeds, final_mask, boxes_embeds, boxes_mask
+            )
+
+        if masks is not None and self.mask_encoder is not None:
+            masks_embed, masks_mask = self._encode_masks(
+                masks=masks,
+                attn_mask=masks_mask,
+                mask_labels=masks_labels,
+                img_feats=img_feats,
+            )
+            if points.size(0) == boxes.size(0) == 0:
+                return masks_embed, masks_mask
+        bs = final_embeds.shape[1]
+        assert final_mask.shape[0] == bs
+        if self.cls_embed is not None:
+            cls = self.cls_embed.weight.view(1, 1, self.d_model).repeat(1, bs, 1)
+            cls_mask = torch.zeros(
+                bs, 1, dtype=final_mask.dtype, device=final_mask.device
+            )
+            final_embeds, final_mask = concat_padded_sequences(
+                final_embeds, final_mask, cls, cls_mask
+            )
+
+        if self.final_proj is not None:
+            final_embeds = self.norm(self.final_proj(final_embeds))
+
+        if self.encode is not None:
+            for lay in self.encode:
+                final_embeds = activation_ckpt_wrapper(lay)(
+                    tgt=final_embeds,
+                    memory=seq_first_img_feats,
+                    tgt_key_padding_mask=final_mask,
+                    pos=seq_first_img_pos_embeds,
+                    act_ckpt_enable=self.training and self.use_act_ckpt,
+                )
+            final_embeds = self.encode_norm(final_embeds)
+        # Finally, concat mask embeddings if any
+        if masks is not None and self.mask_encoder is not None:
+            final_embeds, final_mask = concat_padded_sequences(
+                final_embeds, final_mask, masks_embed, masks_mask
+            )
+        return final_embeds, final_mask

sam3/model/io_utils.py

@@ -0,0 +1,709 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import contextlib
+import os
+import queue
+import re
+import time
+from threading import Condition, get_ident, Lock, Thread
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+import torchvision.transforms.functional as TF
+
+from PIL import Image
+
+from sam3.logger import get_logger
+from tqdm import tqdm
+
+logger = get_logger(__name__)
+
+IS_MAIN_PROCESS = os.getenv("IS_MAIN_PROCESS", "1") == "1"
+RANK = int(os.getenv("RANK", "0"))
+
+IMAGE_EXTS = [".jpg", ".jpeg", ".png", ".bmp", ".tiff", ".webp"]
+VIDEO_EXTS = [".mp4", ".mov", ".avi", ".mkv", ".webm"]
+
+
+def load_resource_as_video_frames(
+    resource_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.5, 0.5, 0.5),
+    img_std=(0.5, 0.5, 0.5),
+    async_loading_frames=False,
+    video_loader_type="cv2",
+):
+    """
+    Load video frames from either a video or an image (as a single-frame video).
+    Alternatively, if input is a list of PIL images, convert its format
+    """
+    if isinstance(resource_path, list):
+        img_mean = torch.tensor(img_mean, dtype=torch.float16)[:, None, None]
+        img_std = torch.tensor(img_std, dtype=torch.float16)[:, None, None]
+        assert all(isinstance(img_pil, Image.Image) for img_pil in resource_path)
+        assert len(resource_path) is not None
+        orig_height, orig_width = resource_path[0].size
+        orig_height, orig_width = (
+            orig_width,
+            orig_height,
+        )  # For some reason, this method returns these swapped
+        images = []
+        for img_pil in resource_path:
+            img_np = np.array(img_pil.convert("RGB").resize((image_size, image_size)))
+            assert img_np.dtype == np.uint8, "np.uint8 is expected for JPEG images"
+            img_np = img_np / 255.0
+            img = torch.from_numpy(img_np).permute(2, 0, 1)
+            # float16 precision should be sufficient for image tensor storage
+            img = img.to(dtype=torch.float16)
+            # normalize by mean and std
+            img -= img_mean
+            img /= img_std
+            images.append(img)
+        images = torch.stack(images)
+        if not offload_video_to_cpu:
+            images = images.cuda()
+        return images, orig_height, orig_width
+
+    is_image = (
+        isinstance(resource_path, str)
+        and os.path.splitext(resource_path)[-1].lower() in IMAGE_EXTS
+    )
+    if is_image:
+        return load_image_as_single_frame_video(
+            image_path=resource_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+        )
+    else:
+        return load_video_frames(
+            video_path=resource_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+            async_loading_frames=async_loading_frames,
+            video_loader_type=video_loader_type,
+        )
+
+
+def load_image_as_single_frame_video(
+    image_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.5, 0.5, 0.5),
+    img_std=(0.5, 0.5, 0.5),
+):
+    """Load an image as a single-frame video."""
+    images, image_height, image_width = _load_img_as_tensor(image_path, image_size)
+    images = images.unsqueeze(0).half()
+
+    img_mean = torch.tensor(img_mean, dtype=torch.float16)[:, None, None]
+    img_std = torch.tensor(img_std, dtype=torch.float16)[:, None, None]
+    if not offload_video_to_cpu:
+        images = images.cuda()
+        img_mean = img_mean.cuda()
+        img_std = img_std.cuda()
+    # normalize by mean and std
+    images -= img_mean
+    images /= img_std
+    return images, image_height, image_width
+
+
+def load_video_frames(
+    video_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.5, 0.5, 0.5),
+    img_std=(0.5, 0.5, 0.5),
+    async_loading_frames=False,
+    video_loader_type="cv2",
+):
+    """
+    Load the video frames from video_path. The frames are resized to image_size as in
+    the model and are loaded to GPU if offload_video_to_cpu=False. This is used by the demo.
+    """
+    assert isinstance(video_path, str)
+    if video_path.startswith("<load-dummy-video"):
+        # Check for pattern <load-dummy-video-N> where N is an integer
+        match = re.match(r"<load-dummy-video-(\d+)>", video_path)
+        num_frames = int(match.group(1)) if match else 60
+        return load_dummy_video(image_size, offload_video_to_cpu, num_frames=num_frames)
+    elif os.path.isdir(video_path):
+        return load_video_frames_from_image_folder(
+            image_folder=video_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+            async_loading_frames=async_loading_frames,
+        )
+    elif os.path.splitext(video_path)[-1].lower() in VIDEO_EXTS:
+        return load_video_frames_from_video_file(
+            video_path=video_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+            async_loading_frames=async_loading_frames,
+            video_loader_type=video_loader_type,
+        )
+    else:
+        raise NotImplementedError("Only video files and image folders are supported")
+
+
+def load_video_frames_from_image_folder(
+    image_folder,
+    image_size,
+    offload_video_to_cpu,
+    img_mean,
+    img_std,
+    async_loading_frames,
+):
+    """
+    Load the video frames from a directory of image files ("<frame_index>.<img_ext>" format)
+    """
+    frame_names = [
+        p
+        for p in os.listdir(image_folder)
+        if os.path.splitext(p)[-1].lower() in IMAGE_EXTS
+    ]
+    try:
+        frame_names.sort(key=lambda p: int(os.path.splitext(p)[0]))
+    except ValueError:
+        # fallback to lexicographic sort if the format is not "<frame_index>.<img_ext>"
+        logger.warning(
+            f'frame names are not in "<frame_index>.<img_ext>" format: {frame_names[:5]=}, '
+            f"falling back to lexicographic sort."
+        )
+        frame_names.sort()
+    num_frames = len(frame_names)
+    if num_frames == 0:
+        raise RuntimeError(f"no images found in {image_folder}")
+    img_paths = [os.path.join(image_folder, frame_name) for frame_name in frame_names]
+    img_mean = torch.tensor(img_mean, dtype=torch.float16)[:, None, None]
+    img_std = torch.tensor(img_std, dtype=torch.float16)[:, None, None]
+
+    if async_loading_frames:
+        lazy_images = AsyncImageFrameLoader(
+            img_paths, image_size, offload_video_to_cpu, img_mean, img_std
+        )
+        return lazy_images, lazy_images.video_height, lazy_images.video_width
+
+    # float16 precision should be sufficient for image tensor storage
+    images = torch.zeros(num_frames, 3, image_size, image_size, dtype=torch.float16)
+    video_height, video_width = None, None
+    for n, img_path in enumerate(
+        tqdm(img_paths, desc=f"frame loading (image folder) [rank={RANK}]")
+    ):
+        images[n], video_height, video_width = _load_img_as_tensor(img_path, image_size)
+    if not offload_video_to_cpu:
+        images = images.cuda()
+        img_mean = img_mean.cuda()
+        img_std = img_std.cuda()
+    # normalize by mean and std
+    images -= img_mean
+    images /= img_std
+    return images, video_height, video_width
+
+
+def load_video_frames_from_video_file(
+    video_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean,
+    img_std,
+    async_loading_frames,
+    gpu_acceleration=False,
+    gpu_device=None,
+    video_loader_type="cv2",
+):
+    """Load the video frames from a video file."""
+    if video_loader_type == "cv2":
+        return load_video_frames_from_video_file_using_cv2(
+            video_path=video_path,
+            image_size=image_size,
+            img_mean=img_mean,
+            img_std=img_std,
+            offload_video_to_cpu=offload_video_to_cpu,
+        )
+    elif video_loader_type == "torchcodec":
+        logger.info("Using torchcodec to load video file")
+        lazy_images = AsyncVideoFileLoaderWithTorchCodec(
+            video_path=video_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+            gpu_acceleration=gpu_acceleration,
+            gpu_device=gpu_device,
+        )
+        # The `AsyncVideoFileLoaderWithTorchCodec` class always loads the videos asynchronously,
+        # so we just wait for its loading thread to finish if async_loading_frames=False.
+        if not async_loading_frames:
+            async_thread = lazy_images.thread
+            if async_thread is not None:
+                async_thread.join()
+        return lazy_images, lazy_images.video_height, lazy_images.video_width
+    else:
+        raise RuntimeError("video_loader_type must be either 'cv2' or 'torchcodec'")
+
+
+def load_video_frames_from_video_file_using_cv2(
+    video_path: str,
+    image_size: int,
+    img_mean: tuple = (0.5, 0.5, 0.5),
+    img_std: tuple = (0.5, 0.5, 0.5),
+    offload_video_to_cpu: bool = False,
+) -> torch.Tensor:
+    """
+    Load video from path, convert to normalized tensor with specified preprocessing
+
+    Args:
+        video_path: Path to video file
+        image_size: Target size for square frames (height and width)
+        img_mean: Normalization mean (RGB)
+        img_std: Normalization standard deviation (RGB)
+
+    Returns:
+        torch.Tensor: Preprocessed video tensor in shape (T, C, H, W) with float16 dtype
+    """
+    import cv2  # delay OpenCV import to avoid unnecessary dependency
+
+    # Initialize video capture
+    cap = cv2.VideoCapture(video_path)
+    if not cap.isOpened():
+        raise ValueError(f"Could not open video: {video_path}")
+
+    original_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    original_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    num_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    num_frames = num_frames if num_frames > 0 else None
+
+    frames = []
+    pbar = tqdm(desc=f"frame loading (OpenCV) [rank={RANK}]", total=num_frames)
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            break
+
+        # Convert BGR to RGB and resize
+        frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+        frame_resized = cv2.resize(
+            frame_rgb, (image_size, image_size), interpolation=cv2.INTER_CUBIC
+        )
+        frames.append(frame_resized)
+        pbar.update(1)
+    cap.release()
+    pbar.close()
+
+    # Convert to tensor
+    frames_np = np.stack(frames, axis=0).astype(np.float32)  # (T, H, W, C)
+    video_tensor = torch.from_numpy(frames_np).permute(0, 3, 1, 2)  # (T, C, H, W)
+
+    img_mean = torch.tensor(img_mean, dtype=torch.float16).view(1, 3, 1, 1)
+    img_std = torch.tensor(img_std, dtype=torch.float16).view(1, 3, 1, 1)
+    if not offload_video_to_cpu:
+        video_tensor = video_tensor.cuda()
+        img_mean = img_mean.cuda()
+        img_std = img_std.cuda()
+    # normalize by mean and std
+    video_tensor -= img_mean
+    video_tensor /= img_std
+    return video_tensor, original_height, original_width
+
+
+def load_dummy_video(image_size, offload_video_to_cpu, num_frames=60):
+    """
+    Load a dummy video with random frames for testing and compilation warmup purposes.
+    """
+    video_height, video_width = 480, 640  # dummy original video sizes
+    images = torch.randn(num_frames, 3, image_size, image_size, dtype=torch.float16)
+    if not offload_video_to_cpu:
+        images = images.cuda()
+    return images, video_height, video_width
+
+
+def _load_img_as_tensor(img_path, image_size):
+    """Load and resize an image and convert it into a PyTorch tensor."""
+    img = Image.open(img_path).convert("RGB")
+    orig_width, orig_height = img.width, img.height
+    img = TF.resize(img, size=(image_size, image_size))
+    img = TF.to_tensor(img)
+    return img, orig_height, orig_width
+
+
+class AsyncImageFrameLoader:
+    """
+    A list of video frames to be load asynchronously without blocking session start.
+    """
+
+    def __init__(self, img_paths, image_size, offload_video_to_cpu, img_mean, img_std):
+        self.img_paths = img_paths
+        self.image_size = image_size
+        self.offload_video_to_cpu = offload_video_to_cpu
+        self.img_mean = img_mean
+        self.img_std = img_std
+        # items in `self._images` will be loaded asynchronously
+        self.images = [None] * len(img_paths)
+        # catch and raise any exceptions in the async loading thread
+        self.exception = None
+        # video_height and video_width be filled when loading the first image
+        self.video_height = None
+        self.video_width = None
+
+        # load the first frame to fill video_height and video_width and also
+        # to cache it (since it's most likely where the user will click)
+        self.__getitem__(0)
+
+        # load the rest of frames asynchronously without blocking the session start
+        def _load_frames():
+            try:
+                for n in tqdm(
+                    range(len(self.images)),
+                    desc=f"frame loading (image folder) [rank={RANK}]",
+                ):
+                    self.__getitem__(n)
+            except Exception as e:
+                self.exception = e
+
+        self.thread = Thread(target=_load_frames, daemon=True)
+        self.thread.start()
+
+    def __getitem__(self, index):
+        if self.exception is not None:
+            raise RuntimeError("Failure in frame loading thread") from self.exception
+
+        img = self.images[index]
+        if img is not None:
+            return img
+
+        img, video_height, video_width = _load_img_as_tensor(
+            self.img_paths[index], self.image_size
+        )
+        self.video_height = video_height
+        self.video_width = video_width
+        # float16 precision should be sufficient for image tensor storage
+        img = img.to(dtype=torch.float16)
+        # normalize by mean and std
+        img -= self.img_mean
+        img /= self.img_std
+        if not self.offload_video_to_cpu:
+            img = img.cuda()
+        self.images[index] = img
+        return img
+
+    def __len__(self):
+        return len(self.images)
+
+
+class TorchCodecDecoder:
+    """
+    A wrapper to support GPU device and num_threads in TorchCodec decoder,
+    which are not supported by `torchcodec.decoders.SimpleVideoDecoder` yet.
+    """
+
+    def __init__(self, source, dimension_order="NCHW", device="cpu", num_threads=1):
+        from torchcodec import _core as core
+
+        self._source = source  # hold a reference to the source to prevent it from GC
+        if isinstance(source, str):
+            self._decoder = core.create_from_file(source, "exact")
+        elif isinstance(source, bytes):
+            self._decoder = core.create_from_bytes(source, "exact")
+        else:
+            raise TypeError(f"Unknown source type: {type(source)}.")
+        assert dimension_order in ("NCHW", "NHWC")
+
+        device_string = str(device)
+        core.scan_all_streams_to_update_metadata(self._decoder)
+        core.add_video_stream(
+            self._decoder,
+            dimension_order=dimension_order,
+            device=device_string,
+            num_threads=(1 if "cuda" in device_string else num_threads),
+        )
+        video_metadata = core.get_container_metadata(self._decoder)
+        best_stream_index = video_metadata.best_video_stream_index
+        assert best_stream_index is not None
+        self.metadata = video_metadata.streams[best_stream_index]
+        assert self.metadata.num_frames_from_content is not None
+        self._num_frames = self.metadata.num_frames_from_content
+
+    def __len__(self) -> int:
+        return self._num_frames
+
+    def __getitem__(self, key: int):
+        from torchcodec import _core as core
+
+        if key < 0:
+            key += self._num_frames
+        if key >= self._num_frames or key < 0:
+            raise IndexError(
+                f"Index {key} is out of bounds; length is {self._num_frames}"
+            )
+        frame_data, *_ = core.get_frame_at_index(
+            self._decoder,
+            frame_index=key,
+        )
+        return frame_data
+
+
+class FIFOLock:
+    """A lock that ensures FIFO ordering of lock acquisitions."""
+
+    def __init__(self):
+        self._lock = Lock()
+        self._waiters = queue.Queue()
+        self._condition = Condition()
+
+    def acquire(self):
+        ident = get_ident()
+        with self._condition:
+            self._waiters.put(ident)
+            while self._waiters.queue[0] != ident or not self._lock.acquire(
+                blocking=False
+            ):
+                self._condition.wait()
+                # got the lock and it's our turn
+
+    def release(self):
+        with self._condition:
+            self._lock.release()
+            self._waiters.get()
+            self._condition.notify_all()
+
+    def __enter__(self):
+        self.acquire()
+
+    def __exit__(self, t, v, tb):
+        self.release()
+
+
+class AsyncVideoFileLoaderWithTorchCodec:
+    """
+    Loading frames from video files asynchronously without blocking session start.
+
+    Unlike `AsyncVideoFileLoader`, this class uses PyTorch's offical TorchCodec library
+    for video decoding, which is more efficient and supports more video formats.
+    """
+
+    def __init__(
+        self,
+        video_path,
+        image_size,
+        offload_video_to_cpu,
+        img_mean,
+        img_std,
+        gpu_acceleration=True,
+        gpu_device=None,
+        use_rand_seek_in_loading=False,
+    ):
+        # Check and possibly infer the output device (and also get its GPU id when applicable)
+        assert gpu_device is None or gpu_device.type == "cuda"
+        gpu_id = (
+            gpu_device.index
+            if gpu_device is not None and gpu_device.index is not None
+            else torch.cuda.current_device()
+        )
+        if offload_video_to_cpu:
+            out_device = torch.device("cpu")
+        else:
+            out_device = torch.device("cuda") if gpu_device is None else gpu_device
+        self.out_device = out_device
+        self.gpu_acceleration = gpu_acceleration
+        self.gpu_id = gpu_id
+        self.image_size = image_size
+        self.offload_video_to_cpu = offload_video_to_cpu
+        if not isinstance(img_mean, torch.Tensor):
+            img_mean = torch.tensor(img_mean, dtype=torch.float16)[:, None, None]
+        self.img_mean = img_mean
+        if not isinstance(img_std, torch.Tensor):
+            img_std = torch.tensor(img_std, dtype=torch.float16)[:, None, None]
+        self.img_std = img_std
+
+        if gpu_acceleration:
+            self.img_mean = self.img_mean.to(f"cuda:{self.gpu_id}")
+            self.img_std = self.img_std.to(f"cuda:{self.gpu_id}")
+            decoder_option = {"device": f"cuda:{self.gpu_id}"}
+        else:
+            self.img_mean = self.img_mean.cpu()
+            self.img_std = self.img_std.cpu()
+            decoder_option = {"num_threads": 1}  # use a single thread to save memory
+
+        self.rank = int(os.environ.get("RANK", "0"))
+        self.world_size = int(os.environ.get("WORLD_SIZE", "1"))
+        self.async_reader = TorchCodecDecoder(video_path, **decoder_option)
+
+        # `num_frames_from_content` is the true number of frames in the video content
+        # from the scan operation (rather than from the metadata, which could be wrong)
+        self.num_frames = self.async_reader.metadata.num_frames_from_content
+        self.video_height = self.async_reader.metadata.height
+        self.video_width = self.async_reader.metadata.width
+
+        # items in `self._images` will be loaded asynchronously
+        self.images_loaded = [False] * self.num_frames
+        self.images = torch.zeros(
+            self.num_frames,
+            3,
+            self.image_size,
+            self.image_size,
+            dtype=torch.float16,
+            device=self.out_device,
+        )
+        # catch and raise any exceptions in the async loading thread
+        self.exception = None
+        self.use_rand_seek_in_loading = use_rand_seek_in_loading
+        self.rand_seek_idx_queue = queue.Queue()
+        # use a lock to avoid race condition between concurrent access to torchcodec
+        # libs (which are not thread-safe); the lock is replaced with a nullcontext
+        # when the video is fully loaded
+        self.torchcodec_access_lock = FIFOLock()
+        self._start_video_loading()
+
+    def _load_one_frame(self, idx):
+        frame_resized = self._transform_frame(self.async_reader[idx])
+        return frame_resized
+
+    @torch.inference_mode()
+    def _start_video_loading(self):
+        desc = f"frame loading (TorchCodec w/ {'GPU' if self.gpu_acceleration else 'CPU'}) [rank={RANK}]"
+        pbar = tqdm(desc=desc, total=self.num_frames)
+        self.num_loaded_frames = 0
+        # load the first frame synchronously to cache it before the session is opened
+        idx = self.num_loaded_frames
+        self.images[idx] = self._load_one_frame(idx)
+        self.images_loaded[idx] = True
+        self.num_loaded_frames += 1
+        pbar.update(n=1)
+        self.all_frames_loaded = self.num_loaded_frames == self.num_frames
+
+        # load the frames asynchronously without blocking the session start
+        def _load_frames():
+            finished = self.all_frames_loaded
+            chunk_size = 16
+            while not finished:
+                # asynchronously load `chunk_size` frames each time we acquire the lock
+                with self.torchcodec_access_lock, torch.inference_mode():
+                    for _ in range(chunk_size):
+                        try:
+                            idx = self.num_loaded_frames
+                            self.images[idx] = self._load_one_frame(idx)
+                            self.images_loaded[idx] = True
+                            self.num_loaded_frames += 1
+                            pbar.update(n=1)
+                            if self.num_loaded_frames >= self.num_frames:
+                                finished = True
+                                break
+                        except Exception as e:
+                            self.exception = e
+                            raise
+
+                    # also read the frame that is being randomly seeked to
+                    while True:
+                        try:
+                            idx = self.rand_seek_idx_queue.get_nowait()
+                            if not self.images_loaded[idx]:
+                                self.images[idx] = self._load_one_frame(idx)
+                                self.images_loaded[idx] = True
+                        except queue.Empty:
+                            break
+                        except Exception as e:
+                            self.exception = e
+                            raise
+
+            # finished -- check whether we have loaded the total number of frames
+            if self.num_loaded_frames != self.num_frames:
+                raise RuntimeError(
+                    f"There are {self.num_frames} frames in the video, but only "
+                    f"{self.num_loaded_frames} frames can be loaded successfully."
+                )
+            else:
+                self.all_frames_loaded = True
+                pbar.close()
+                with self.torchcodec_access_lock:
+                    import gc
+
+                    # all frames have been loaded, so we can release the readers and free their memory
+                    # also remove pbar and thread (which shouldn't be a part of session saving)
+                    reader = self.async_reader
+                    if reader is not None:
+                        reader._source = None
+                    self.async_reader = None
+                    self.pbar = None
+                    self.thread = None
+                    self.rand_seek_idx_queue = None
+                    gc.collect()
+                # remove the lock (replace it with nullcontext) when the video is fully loaded
+                self.torchcodec_access_lock = contextlib.nullcontext()
+
+        self.thread = Thread(target=_load_frames, daemon=True)
+        self.thread.start()
+
+    def _transform_frame(self, frame):
+        frame = frame.clone()  # make a copy to avoid modifying the original frame bytes
+        frame = frame.float()  # convert to float32 before interpolation
+        frame_resized = F.interpolate(
+            frame[None, :],
+            size=(self.image_size, self.image_size),
+            mode="bicubic",
+            align_corners=False,
+        )[0]
+        # float16 precision should be sufficient for image tensor storage
+        frame_resized = frame_resized.half()  # uint8 -> float16
+        frame_resized /= 255
+        frame_resized -= self.img_mean
+        frame_resized /= self.img_std
+        if self.offload_video_to_cpu:
+            frame_resized = frame_resized.cpu()
+        elif frame_resized.device != self.out_device:
+            frame_resized = frame_resized.to(device=self.out_device, non_blocking=True)
+        return frame_resized
+
+    def __getitem__(self, index):
+        if self.exception is not None:
+            raise RuntimeError("Failure in frame loading thread") from self.exception
+
+        max_tries = 1200
+        for _ in range(max_tries):
+            # use a lock to avoid race condition between concurrent access to torchcodec
+            # libs (which are not thread-safe); the lock is replaced with a nullcontext
+            # when the video is fully loaded
+            with self.torchcodec_access_lock:
+                if self.images_loaded[index]:
+                    return self.images[index]
+
+                if self.use_rand_seek_in_loading:
+                    # async loading hasn't reached this frame yet, so we load this frame individually
+                    # (it will be loaded by in _load_frames thread and added to self.images[index])
+                    self.rand_seek_idx_queue.put(index)
+
+            time.sleep(0.1)
+
+        raise RuntimeError(f"Failed to load frame {index} after {max_tries} tries")
+
+    def __len__(self):
+        return len(self.images)
+
+    def __getstate__(self):
+        """
+        Remove a few attributes during pickling, so that this async video loader can be
+        saved and loaded as a part of the model session.
+        """
+        # wait for async video loading to finish before pickling
+        async_thread = self.thread
+        if async_thread is not None:
+            async_thread.join()
+        # release a few objects that cannot be pickled
+        reader = self.async_reader
+        if reader is not None:
+            reader._source = None
+        self.async_reader = None
+        self.pbar = None
+        self.thread = None
+        self.rand_seek_idx_queue = None
+        self.torchcodec_access_lock = contextlib.nullcontext()
+        return self.__dict__.copy()

sam3/model/maskformer_segmentation.py

@@ -0,0 +1,323 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import math
+from typing import Dict, List, Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+
+from .model_misc import MLP
+
+
+class LinearPresenceHead(nn.Sequential):
+    def __init__(self, d_model):
+        # a hack to make `LinearPresenceHead` compatible with old checkpoints
+        super().__init__(nn.Identity(), nn.Identity(), nn.Linear(d_model, 1))
+
+    def forward(self, hs, prompt, prompt_mask):
+        return super().forward(hs)
+
+
+class MaskPredictor(nn.Module):
+    def __init__(self, hidden_dim, mask_dim):
+        super().__init__()
+        self.mask_embed = MLP(hidden_dim, hidden_dim, mask_dim, 3)
+
+    def forward(self, obj_queries, pixel_embed):
+        if len(obj_queries.shape) == 3:
+            if pixel_embed.ndim == 3:
+                # batch size was omitted
+                mask_preds = torch.einsum(
+                    "bqc,chw->bqhw", self.mask_embed(obj_queries), pixel_embed
+                )
+            else:
+                mask_preds = torch.einsum(
+                    "bqc,bchw->bqhw", self.mask_embed(obj_queries), pixel_embed
+                )
+        else:
+            # Assumed to have aux masks
+            if pixel_embed.ndim == 3:
+                # batch size was omitted
+                mask_preds = torch.einsum(
+                    "lbqc,chw->lbqhw", self.mask_embed(obj_queries), pixel_embed
+                )
+            else:
+                mask_preds = torch.einsum(
+                    "lbqc,bchw->lbqhw", self.mask_embed(obj_queries), pixel_embed
+                )
+
+        return mask_preds
+
+
+class SegmentationHead(nn.Module):
+    def __init__(
+        self,
+        hidden_dim,
+        upsampling_stages,
+        use_encoder_inputs=False,
+        aux_masks=False,
+        no_dec=False,
+        pixel_decoder=None,
+        act_ckpt=False,
+        shared_conv=False,
+        compile_mode_pixel_decoder=None,
+    ):
+        super().__init__()
+        self.use_encoder_inputs = use_encoder_inputs
+        self.aux_masks = aux_masks
+        if pixel_decoder is not None:
+            self.pixel_decoder = pixel_decoder
+        else:
+            self.pixel_decoder = PixelDecoder(
+                hidden_dim,
+                upsampling_stages,
+                shared_conv=shared_conv,
+                compile_mode=compile_mode_pixel_decoder,
+            )
+        self.no_dec = no_dec
+        if no_dec:
+            self.mask_predictor = nn.Conv2d(
+                hidden_dim, 1, kernel_size=3, stride=1, padding=1
+            )
+        else:
+            self.mask_predictor = MaskPredictor(hidden_dim, mask_dim=hidden_dim)
+
+        self.act_ckpt = act_ckpt
+
+        # used to update the output dictionary
+        self.instance_keys = ["pred_masks"]
+
+    @property
+    def device(self):
+        self._device = getattr(self, "_device", None) or next(self.parameters()).device
+        return self._device
+
+    def to(self, *args, **kwargs):
+        # clear cached _device in case the model is moved to a different device
+        self._device = None
+        return super().to(*args, **kwargs)
+
+    def _embed_pixels(
+        self,
+        backbone_feats: List[torch.Tensor],
+        image_ids,
+        encoder_hidden_states,
+    ) -> torch.Tensor:
+        feature_device = backbone_feats[0].device  # features could be on CPU
+        model_device = self.device
+        image_ids_ = image_ids.to(feature_device)
+        if self.use_encoder_inputs:
+            if backbone_feats[0].shape[0] > 1:
+                # For bs > 1, we construct the per query backbone features
+                backbone_visual_feats = []
+                for feat in backbone_feats:
+                    # Copy the img features per query (pixel decoder won't share img feats)
+                    backbone_visual_feats.append(feat[image_ids_, ...].to(model_device))
+            else:
+                # Bs=1, we rely on broadcasting for query-based processing
+                backbone_visual_feats = [bb_feat.clone() for bb_feat in backbone_feats]
+            # Extract visual embeddings
+            encoder_hidden_states = encoder_hidden_states.permute(1, 2, 0)
+            spatial_dim = math.prod(backbone_feats[-1].shape[-2:])
+            encoder_visual_embed = encoder_hidden_states[..., :spatial_dim].reshape(
+                -1, *backbone_feats[-1].shape[1:]
+            )
+
+            backbone_visual_feats[-1] = encoder_visual_embed
+            if self.act_ckpt:
+                pixel_embed = checkpoint.checkpoint(
+                    self.pixel_decoder, backbone_visual_feats, use_reentrant=False
+                )
+            else:
+                pixel_embed = self.pixel_decoder(backbone_visual_feats)
+        else:
+            backbone_feats = [x.to(model_device) for x in backbone_feats]
+            pixel_embed = self.pixel_decoder(backbone_feats)
+            if pixel_embed.shape[0] == 1:
+                # For batch_size=1 training, we can avoid the indexing to save memory
+                pixel_embed = pixel_embed.squeeze(0)
+            else:
+                pixel_embed = pixel_embed[image_ids, ...]
+        return pixel_embed
+
+    def forward(
+        self,
+        backbone_feats: List[torch.Tensor],
+        obj_queries: torch.Tensor,
+        image_ids,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Dict[str, torch.Tensor]:
+        if self.use_encoder_inputs:
+            assert encoder_hidden_states is not None
+
+        pixel_embed = self._embed_pixels(
+            backbone_feats=backbone_feats,
+            image_ids=image_ids,
+            encoder_hidden_states=encoder_hidden_states,
+        )
+
+        if self.no_dec:
+            mask_pred = self.mask_predictor(pixel_embed)
+        elif self.aux_masks:
+            mask_pred = self.mask_predictor(obj_queries, pixel_embed)
+        else:
+            mask_pred = self.mask_predictor(obj_queries[-1], pixel_embed)
+
+        return {"pred_masks": mask_pred}
+
+
+class PixelDecoder(nn.Module):
+    def __init__(
+        self,
+        hidden_dim,
+        num_upsampling_stages,
+        interpolation_mode="nearest",
+        shared_conv=False,
+        compile_mode=None,
+    ):
+        super().__init__()
+        self.hidden_dim = hidden_dim
+        self.num_upsampling_stages = num_upsampling_stages
+        self.interpolation_mode = interpolation_mode
+        conv_layers = []
+        norms = []
+        num_convs = 1 if shared_conv else num_upsampling_stages
+        for _ in range(num_convs):
+            conv_layers.append(nn.Conv2d(self.hidden_dim, self.hidden_dim, 3, 1, 1))
+            norms.append(nn.GroupNorm(8, self.hidden_dim))
+
+        self.conv_layers = nn.ModuleList(conv_layers)
+        self.norms = nn.ModuleList(norms)
+        self.shared_conv = shared_conv
+        self.out_dim = self.conv_layers[-1].out_channels
+        if compile_mode is not None:
+            self.forward = torch.compile(
+                self.forward, mode=compile_mode, dynamic=True, fullgraph=True
+            )
+            # Needed to make checkpointing happy. But we don't know if the module is checkpointed, so we disable it by default.
+            torch._dynamo.config.optimize_ddp = False
+
+    def forward(self, backbone_feats: List[torch.Tensor]):
+        # Assumes backbone features are already projected (C == hidden dim)
+
+        prev_fpn = backbone_feats[-1]
+        fpn_feats = backbone_feats[:-1]
+        for layer_idx, bb_feat in enumerate(fpn_feats[::-1]):
+            curr_fpn = bb_feat
+            prev_fpn = curr_fpn + F.interpolate(
+                prev_fpn, size=curr_fpn.shape[-2:], mode=self.interpolation_mode
+            )
+            if self.shared_conv:
+                # only one conv layer
+                layer_idx = 0
+            prev_fpn = self.conv_layers[layer_idx](prev_fpn)
+            prev_fpn = F.relu(self.norms[layer_idx](prev_fpn))
+
+        return prev_fpn
+
+
+class UniversalSegmentationHead(SegmentationHead):
+    """This module handles semantic+instance segmentation"""
+
+    def __init__(
+        self,
+        hidden_dim,
+        upsampling_stages,
+        pixel_decoder,
+        aux_masks=False,
+        no_dec=False,
+        act_ckpt=False,
+        presence_head: bool = False,
+        dot_product_scorer=None,
+        cross_attend_prompt=None,
+    ):
+        super().__init__(
+            hidden_dim=hidden_dim,
+            upsampling_stages=upsampling_stages,
+            use_encoder_inputs=True,
+            aux_masks=aux_masks,
+            no_dec=no_dec,
+            pixel_decoder=pixel_decoder,
+            act_ckpt=act_ckpt,
+        )
+        self.d_model = hidden_dim
+
+        if dot_product_scorer is not None:
+            assert presence_head, "Specifying a dot product scorer without a presence head is likely a mistake"
+
+        self.presence_head = None
+        if presence_head:
+            self.presence_head = (
+                dot_product_scorer
+                if dot_product_scorer is not None
+                else LinearPresenceHead(self.d_model)
+            )
+
+        self.cross_attend_prompt = cross_attend_prompt
+        if self.cross_attend_prompt is not None:
+            self.cross_attn_norm = nn.LayerNorm(self.d_model)
+
+        self.semantic_seg_head = nn.Conv2d(self.pixel_decoder.out_dim, 1, kernel_size=1)
+        self.instance_seg_head = nn.Conv2d(
+            self.pixel_decoder.out_dim, self.d_model, kernel_size=1
+        )
+
+    def forward(
+        self,
+        backbone_feats: List[torch.Tensor],
+        obj_queries: torch.Tensor,
+        image_ids,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        prompt: Optional[torch.Tensor] = None,
+        prompt_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Dict[str, Optional[torch.Tensor]]:
+        assert encoder_hidden_states is not None
+        bs = encoder_hidden_states.shape[1]
+
+        if self.cross_attend_prompt is not None:
+            tgt2 = self.cross_attn_norm(encoder_hidden_states)
+            tgt2 = self.cross_attend_prompt(
+                query=tgt2,
+                key=prompt,
+                value=prompt,
+                key_padding_mask=prompt_mask,
+            )[0]
+            encoder_hidden_states = tgt2 + encoder_hidden_states
+
+        presence_logit = None
+        if self.presence_head is not None:
+            pooled_enc = encoder_hidden_states.mean(0)
+            presence_logit = (
+                self.presence_head(
+                    pooled_enc.view(1, bs, 1, self.d_model),
+                    prompt=prompt,
+                    prompt_mask=prompt_mask,
+                )
+                .squeeze(0)
+                .squeeze(1)
+            )
+
+        pixel_embed = self._embed_pixels(
+            backbone_feats=backbone_feats,
+            image_ids=image_ids,
+            encoder_hidden_states=encoder_hidden_states,
+        )
+
+        instance_embeds = self.instance_seg_head(pixel_embed)
+
+        if self.no_dec:
+            mask_pred = self.mask_predictor(instance_embeds)
+        elif self.aux_masks:
+            mask_pred = self.mask_predictor(obj_queries, instance_embeds)
+        else:
+            mask_pred = self.mask_predictor(obj_queries[-1], instance_embeds)
+
+        return {
+            "pred_masks": mask_pred,
+            "semantic_seg": self.semantic_seg_head(pixel_embed),
+            "presence_logit": presence_logit,
+        }

sam3/model/memory.py

@@ -0,0 +1,201 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import math
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+try:
+    from timm.layers import DropPath
+except ModuleNotFoundError:
+    # compatibility for older timm versions
+    from timm.models.layers import DropPath
+
+from .model_misc import get_clones, LayerNorm2d
+
+
+class SimpleMaskDownSampler(nn.Module):
+    """
+    Progressively downsample a mask by total_stride, each time by stride.
+    Note that LayerNorm is applied per *token*, like in ViT.
+
+    With each downsample (by a factor stride**2), channel capacity increases by the same factor.
+    In the end, we linearly project to embed_dim channels.
+    """
+
+    def __init__(
+        self,
+        embed_dim=256,
+        kernel_size=4,
+        stride=4,
+        padding=0,
+        total_stride=16,
+        activation=nn.GELU,
+        # Option to interpolate the input mask first before downsampling using convs. In that case, the total_stride is assumed to be after interpolation.
+        # If set to input resolution or None, we don't interpolate. We default to None to be safe (for older configs or if not explicitly set)
+        interpol_size=None,
+    ):
+        super().__init__()
+        num_layers = int(math.log2(total_stride) // math.log2(stride))
+        assert stride**num_layers == total_stride
+        self.encoder = nn.Sequential()
+        mask_in_chans, mask_out_chans = 1, 1
+        for _ in range(num_layers):
+            mask_out_chans = mask_in_chans * (stride**2)
+            self.encoder.append(
+                nn.Conv2d(
+                    mask_in_chans,
+                    mask_out_chans,
+                    kernel_size=kernel_size,
+                    stride=stride,
+                    padding=padding,
+                )
+            )
+            self.encoder.append(LayerNorm2d(mask_out_chans))
+            self.encoder.append(activation())
+            mask_in_chans = mask_out_chans
+
+        self.encoder.append(nn.Conv2d(mask_out_chans, embed_dim, kernel_size=1))
+        self.interpol_size = interpol_size
+        if self.interpol_size is not None:
+            assert isinstance(
+                self.interpol_size, (list, tuple)
+            ), f"Unsupported type {type(self.interpol_size)}. Should be a list or tuple."
+            self.interpol_size = list(interpol_size)
+            assert len(self.interpol_size) == 2
+
+    def forward(self, x: torch.Tensor):
+        if self.interpol_size is not None and self.interpol_size != list(x.shape[-2:]):
+            x = F.interpolate(
+                x.float(),
+                size=self.interpol_size,
+                align_corners=False,
+                mode="bilinear",
+                antialias=True,
+            )
+        return self.encoder(x)
+
+
+# Lightly adapted from ConvNext (https://github.com/facebookresearch/ConvNeXt)
+class CXBlock(nn.Module):
+    r"""ConvNeXt Block. There are two equivalent implementations:
+    (1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N, C, H, W)
+    (2) DwConv -> Permute to (N, H, W, C); LayerNorm (channels_last) -> Linear -> GELU -> Linear; Permute back
+    We use (2) as we find it slightly faster in PyTorch
+
+    Args:
+        dim (int): Number of input channels.
+        drop_path (float): Stochastic depth rate. Default: 0.0
+        layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
+    """
+
+    def __init__(
+        self,
+        dim,
+        kernel_size=7,
+        padding=3,
+        drop_path=0.0,
+        layer_scale_init_value=1e-6,
+        use_dwconv=True,
+    ):
+        super().__init__()
+        self.dwconv = nn.Conv2d(
+            dim,
+            dim,
+            kernel_size=kernel_size,
+            padding=padding,
+            groups=dim if use_dwconv else 1,
+        )  # depthwise conv
+        self.norm = LayerNorm2d(dim, eps=1e-6)
+        self.pwconv1 = nn.Linear(
+            dim, 4 * dim
+        )  # pointwise/1x1 convs, implemented with linear layers
+        self.act = nn.GELU()
+        self.pwconv2 = nn.Linear(4 * dim, dim)
+        self.gamma = (
+            nn.Parameter(layer_scale_init_value * torch.ones((dim)), requires_grad=True)
+            if layer_scale_init_value > 0
+            else None
+        )
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+    def forward(self, x):
+        input = x
+        x = self.dwconv(x)
+        x = self.norm(x)
+        x = x.permute(0, 2, 3, 1)  # (N, C, H, W) -> (N, H, W, C)
+        x = self.pwconv1(x)
+        x = self.act(x)
+        x = self.pwconv2(x)
+        if self.gamma is not None:
+            x = self.gamma * x
+        x = x.permute(0, 3, 1, 2)  # (N, H, W, C) -> (N, C, H, W)
+
+        x = input + self.drop_path(x)
+        return x
+
+
+class SimpleFuser(nn.Module):
+    def __init__(self, layer, num_layers, dim=None, input_projection=False):
+        super().__init__()
+        self.proj = nn.Identity()
+        self.layers = get_clones(layer, num_layers)
+
+        if input_projection:
+            assert dim is not None
+            self.proj = nn.Conv2d(dim, dim, kernel_size=1)
+
+    def forward(self, x):
+        # normally x: (N, C, H, W)
+        x = self.proj(x)
+        for layer in self.layers:
+            x = layer(x)
+        return x
+
+
+class SimpleMaskEncoder(nn.Module):
+    def __init__(
+        self,
+        out_dim,
+        mask_downsampler,
+        fuser,
+        position_encoding,
+        in_dim=256,  # in_dim of pix_feats
+    ):
+        super().__init__()
+
+        self.mask_downsampler = mask_downsampler
+
+        self.pix_feat_proj = nn.Conv2d(in_dim, in_dim, kernel_size=1)
+        self.fuser = fuser
+        self.position_encoding = position_encoding
+        self.out_proj = nn.Identity()
+        if out_dim != in_dim:
+            self.out_proj = nn.Conv2d(in_dim, out_dim, kernel_size=1)
+
+    def forward(
+        self,
+        pix_feat: torch.Tensor,
+        masks: torch.Tensor,
+        skip_mask_sigmoid: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        ## Process masks
+        # sigmoid, so that less domain shift from gt masks which are bool
+        if not skip_mask_sigmoid:
+            masks = F.sigmoid(masks)
+        masks = self.mask_downsampler(masks)
+
+        ## Fuse pix_feats and downsampled masks
+        # in case the visual features are on CPU, cast them to CUDA
+        pix_feat = pix_feat.to(masks.device)
+
+        x = self.pix_feat_proj(pix_feat)
+        x = x + masks
+        x = self.fuser(x)
+        x = self.out_proj(x)
+
+        pos = self.position_encoding(x).to(x.dtype)
+
+        return {"vision_features": x, "vision_pos_enc": [pos]}

sam3/model/model_misc.py

@@ -0,0 +1,428 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+"""Various utility models"""
+
+import copy
+import math
+import weakref
+from collections.abc import Iterator
+from contextlib import AbstractContextManager
+from enum import auto, Enum
+from typing import Dict, List, Optional, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn, Tensor
+from typing_extensions import override
+
+
+def inverse_sigmoid(x, eps=1e-3):
+    """
+    The inverse function for sigmoid activation function.
+    Note: It might face numberical issues with fp16 small eps.
+    """
+    x = x.clamp(min=0, max=1)
+    x1 = x.clamp(min=eps)
+    x2 = (1 - x).clamp(min=eps)
+    return torch.log(x1 / x2)
+
+
+class MultiheadAttentionWrapper(nn.MultiheadAttention):
+    def forward(self, *args, **kwargs):
+        kwargs["need_weights"] = False
+        return super().forward(*args, **kwargs)
+
+
+class DotProductScoring(torch.nn.Module):
+    def __init__(
+        self,
+        d_model,
+        d_proj,
+        prompt_mlp=None,
+        clamp_logits=True,
+        clamp_max_val=12.0,
+    ):
+        super().__init__()
+        self.d_proj = d_proj
+        assert isinstance(prompt_mlp, torch.nn.Module) or prompt_mlp is None
+        self.prompt_mlp = prompt_mlp  # an optional MLP projection for prompt
+        self.prompt_proj = torch.nn.Linear(d_model, d_proj)
+        self.hs_proj = torch.nn.Linear(d_model, d_proj)
+        self.scale = float(1.0 / np.sqrt(d_proj))
+        self.clamp_logits = clamp_logits
+        if self.clamp_logits:
+            self.clamp_max_val = clamp_max_val
+
+    def mean_pool_text(self, prompt, prompt_mask):
+        # is_valid has shape (seq, bs, 1), where 1 is valid and 0 is padding
+        is_valid = (~prompt_mask).float().permute(1, 0)[..., None]
+        # num_valid has shape (bs, 1)
+        num_valid = torch.clamp(torch.sum(is_valid, dim=0), min=1.0)
+        # mean pool over all the valid tokens -- pooled_prompt has shape (bs, proj_dim)
+        pooled_prompt = (prompt * is_valid).sum(dim=0) / num_valid
+        return pooled_prompt
+
+    def forward(self, hs, prompt, prompt_mask):
+        # hs has shape (num_layer, bs, num_query, d_model)
+        # prompt has shape (seq, bs, d_model)
+        # prompt_mask has shape (bs, seq), where 1 is valid and 0 is padding
+        assert hs.dim() == 4 and prompt.dim() == 3 and prompt_mask.dim() == 2
+
+        # apply MLP on prompt if specified
+        if self.prompt_mlp is not None:
+            prompt = self.prompt_mlp(prompt)
+
+        # first, get the mean-pooled version of the prompt
+        pooled_prompt = self.mean_pool_text(prompt, prompt_mask)
+
+        # then, project pooled_prompt and hs to d_proj dimensions
+        proj_pooled_prompt = self.prompt_proj(pooled_prompt)  # (bs, d_proj)
+        proj_hs = self.hs_proj(hs)  # (num_layer, bs, num_query, d_proj)
+
+        # finally, get dot-product scores of shape (num_layer, bs, num_query, 1)
+        scores = torch.matmul(proj_hs, proj_pooled_prompt.unsqueeze(-1))
+        scores *= self.scale
+
+        # clamp scores to a max value to avoid numerical issues in loss or matcher
+        if self.clamp_logits:
+            scores.clamp_(min=-self.clamp_max_val, max=self.clamp_max_val)
+
+        return scores
+
+
+class LayerScale(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        init_values: Union[float, Tensor] = 1e-5,
+        inplace: bool = False,
+    ) -> None:
+        super().__init__()
+        self.inplace = inplace
+        self.gamma = nn.Parameter(init_values * torch.ones(dim))
+
+    def forward(self, x: Tensor) -> Tensor:
+        return x.mul_(self.gamma) if self.inplace else x * self.gamma
+
+
+class LayerNorm2d(nn.Module):
+    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(num_channels))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        x = self.weight[:, None, None] * x + self.bias[:, None, None]
+        return x
+
+
+class TransformerWrapper(nn.Module):
+    def __init__(
+        self,
+        encoder,
+        decoder,
+        d_model: int,
+        two_stage_type="none",  # ["none"] only for now
+        pos_enc_at_input_dec=True,
+    ):
+        super().__init__()
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.num_queries = decoder.num_queries if decoder is not None else None
+        self.pos_enc_at_input_dec = pos_enc_at_input_dec
+
+        # for two stage
+        assert two_stage_type in ["none"], "unknown param {} of two_stage_type".format(
+            two_stage_type
+        )
+        self.two_stage_type = two_stage_type
+
+        self._reset_parameters()
+        self.d_model = d_model
+
+    def _reset_parameters(self):
+        for n, p in self.named_parameters():
+            if p.dim() > 1:
+                if (
+                    "box_embed" not in n
+                    and "query_embed" not in n
+                    and "reference_points" not in n
+                ):
+                    nn.init.xavier_uniform_(p)
+
+
+class MLP(nn.Module):
+    """Very simple multi-layer perceptron (also called FFN)"""
+
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        output_dim: int,
+        num_layers: int,
+        dropout: float = 0.0,
+        residual: bool = False,
+        out_norm: Optional[nn.Module] = None,
+    ):
+        super().__init__()
+        self.num_layers = num_layers
+        h = [hidden_dim] * (num_layers - 1)
+        self.layers = nn.ModuleList(
+            nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim])
+        )
+        self.drop = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        # whether to add the output as a residual connection to the input
+        if residual and input_dim != output_dim:
+            raise ValueError("residual is only supported if input_dim == output_dim")
+        self.residual = residual
+        # whether to apply a normalization layer to the output
+        assert isinstance(out_norm, nn.Module) or out_norm is None
+        self.out_norm = out_norm or nn.Identity()
+
+    def forward(self, x):
+        orig_x = x
+        for i, layer in enumerate(self.layers):
+            x = self.drop(F.relu(layer(x))) if i < self.num_layers - 1 else layer(x)
+        if self.residual:
+            x = x + orig_x
+        x = self.out_norm(x)
+        return x
+
+
+def get_clones(module, N):
+    return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
+
+
+def get_clones_seq(module, N):
+    return nn.Sequential(*[copy.deepcopy(module) for i in range(N)])
+
+
+def get_activation_fn(activation):
+    """Return an activation function given a string"""
+    if activation == "relu":
+        return F.relu
+    if activation == "gelu":
+        return F.gelu
+    if activation == "glu":
+        return F.glu
+    raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
+
+
+def get_activation_module(activation):
+    """Return an activation function given a string"""
+    if activation == "relu":
+        return nn.ReLU
+    if activation == "gelu":
+        return nn.GELU
+    if activation == "glu":
+        return nn.GLU
+    raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
+
+
+def get_valid_ratio(mask):
+    _, H, W = mask.shape
+    valid_H = torch.sum(~mask[:, :, 0], 1)
+    valid_W = torch.sum(~mask[:, 0, :], 1)
+    valid_ratio_h = valid_H.float() / H
+    valid_ratio_w = valid_W.float() / W
+    valid_ratio = torch.stack([valid_ratio_w, valid_ratio_h], -1)
+    return valid_ratio
+
+
+def gen_sineembed_for_position(pos_tensor, num_feats=256):
+    assert num_feats % 2 == 0
+    num_feats = num_feats // 2
+    # n_query, bs, _ = pos_tensor.size()
+    # sineembed_tensor = torch.zeros(n_query, bs, 256)
+    scale = 2 * math.pi
+    dim_t = torch.arange(num_feats, dtype=torch.float32, device=pos_tensor.device)
+    dim_t = 10000 ** (2 * (torch.div(dim_t, 2, rounding_mode="floor")) / num_feats)
+    x_embed = pos_tensor[:, :, 0] * scale
+    y_embed = pos_tensor[:, :, 1] * scale
+    pos_x = x_embed[:, :, None] / dim_t
+    pos_y = y_embed[:, :, None] / dim_t
+    pos_x = torch.stack(
+        (pos_x[:, :, 0::2].sin(), pos_x[:, :, 1::2].cos()), dim=3
+    ).flatten(2)
+    pos_y = torch.stack(
+        (pos_y[:, :, 0::2].sin(), pos_y[:, :, 1::2].cos()), dim=3
+    ).flatten(2)
+    if pos_tensor.size(-1) == 2:
+        pos = torch.cat((pos_y, pos_x), dim=2)
+    elif pos_tensor.size(-1) == 4:
+        w_embed = pos_tensor[:, :, 2] * scale
+        pos_w = w_embed[:, :, None] / dim_t
+        pos_w = torch.stack(
+            (pos_w[:, :, 0::2].sin(), pos_w[:, :, 1::2].cos()), dim=3
+        ).flatten(2)
+
+        h_embed = pos_tensor[:, :, 3] * scale
+        pos_h = h_embed[:, :, None] / dim_t
+        pos_h = torch.stack(
+            (pos_h[:, :, 0::2].sin(), pos_h[:, :, 1::2].cos()), dim=3
+        ).flatten(2)
+
+        pos = torch.cat((pos_y, pos_x, pos_w, pos_h), dim=2)
+    else:
+        raise ValueError("Unknown pos_tensor shape(-1):{}".format(pos_tensor.size(-1)))
+    return pos
+
+
+class SAM3Output(list):
+    """
+    A class representing the output of a SAM3 model.
+    It provides an iterable interface that supports different iteration modes, including iterating over all steps per stage,
+    last step per stage, and flattened output.
+    Attributes:
+        output: The output of the SAM3 model, represented as a list of lists.
+        iter_mode: The current iteration mode.
+    Example:
+        >>> output = [[1, 2], [3, 4], [5, 6]]
+        >>> sam3_output = SAM3Output(output)
+        >>> for step in sam3_output:
+        ...     print(step)
+        [1, 2]
+        [3, 4]
+        [5, 6]
+        >>> with SAM3Output.iteration_mode(SAM3Output.IterMode.LAST_STEP_PER_STAGE) as sam3_last_step_out:
+        ...     for step in sam3_last_step_out:
+        ...         print(step)
+        [2]
+        [4]
+        [6]
+        >>> with SAM3Output.iteration_mode(SAM3Output.IterMode.FLATTENED) as sam3_flattened_out:
+        ...     for step in sam3_flattened_out:
+        ...         print(step)
+        1
+        2
+        3
+        4
+        5
+        6
+    """
+
+    class IterMode(Enum):
+        # Defines the type of iterator over ouptuts.
+        ALL_STEPS_PER_STAGE = auto()
+        LAST_STEP_PER_STAGE = auto()
+        FLATTENED = auto()  # Returns each interactivity step as if it is a separate stage (this is used in SAM3Image model)
+
+    def __init__(
+        self,
+        output: List[List[Dict]] = None,
+        iter_mode: IterMode = IterMode.ALL_STEPS_PER_STAGE,
+        loss_stages: Optional[List[int]] = None,
+    ):
+        if output is not None:
+            assert (
+                isinstance(output, list)
+                and len(output) > 0
+                and isinstance(output[0], list)
+            ), "Expected output to be a list of lists"
+            self.output = output
+        else:
+            self.output = []
+        assert isinstance(
+            iter_mode, SAM3Output.IterMode
+        ), f"iter_mode shoulf be of enum type 'SAM3Output.IterMode'. Got {type(iter_mode)}"
+
+        self.iter_mode = iter_mode
+        # We create a weak reference to self to be used in the lambda functions.
+        # This is to avoid cyclic references and let SAM3Output be garabge collected.
+        self_ref = weakref.ref(self)
+        self._mode2iter = {
+            SAM3Output.IterMode.ALL_STEPS_PER_STAGE: lambda: iter(self_ref().output),
+            SAM3Output.IterMode.LAST_STEP_PER_STAGE: lambda: (
+                inner_list[-1] for inner_list in self_ref().output
+            ),
+            SAM3Output.IterMode.FLATTENED: lambda: (
+                element for inner_list in self_ref().output for element in inner_list
+            ),
+        }
+        self.loss_stages = loss_stages
+
+    @override
+    def __iter__(self) -> Iterator:
+        return self._mode2iter[self.iter_mode]()
+
+    def __getitem__(self, index):
+        """
+        Returns the item at the specified index.
+        Args:
+            index (int): The index of the item to return.
+        Returns:
+            list or element: The item at the specified index.
+        """
+        assert isinstance(index, int), f"index should be an integer. Got {type(index)}"
+        if self.iter_mode == SAM3Output.IterMode.ALL_STEPS_PER_STAGE:
+            return self.output[index]
+        elif self.iter_mode == SAM3Output.IterMode.LAST_STEP_PER_STAGE:
+            return self.output[index][-1]
+        elif self.iter_mode == SAM3Output.IterMode.FLATTENED:
+            if index == -1:
+                return self.self.output[-1][-1]
+            else:
+                flattened_output = sum(self.output, [])
+                return flattened_output[index]
+
+    class _IterationMode(AbstractContextManager):
+        """
+        A context manager that temporarily changes the iteration mode of a SAM3Output object.
+        This class is used internally by the SAM3Output.iteration_mode method.
+        """
+
+        def __init__(
+            self, model_output: "SAM3Output", iter_mode: "SAM3Output.IterMode"
+        ):
+            self._model_output = model_output
+            self._orig_iter_mode = model_output.iter_mode
+            self._new_iter_mode = iter_mode
+
+        @override
+        def __enter__(self) -> "SAM3Output":
+            self._model_output.iter_mode = self._new_iter_mode
+            return self._model_output
+
+        @override
+        def __exit__(self, exc_type, exc_value, traceback):
+            self._model_output.iter_mode = self._orig_iter_mode
+            return super().__exit__(exc_type, exc_value, traceback)
+
+    @staticmethod
+    def iteration_mode(
+        model_output: "SAM3Output", iter_mode: IterMode
+    ) -> _IterationMode:
+        """
+        Returns a context manager that allows you to temporarily change the iteration mode of the SAM3Output object.
+        Args:
+            model_output: The SAM3Output object.
+            iter_mode: The new iteration mode.
+        Returns:
+            SAM3Output._IterationMode: A context manager that changes the iteration mode of the SAM3Output object.
+        """
+        return SAM3Output._IterationMode(model_output=model_output, iter_mode=iter_mode)
+
+    def append(self, item: list):
+        assert isinstance(
+            item, list
+        ), f"Only list items are supported. Got {type(item)}"
+        self.output.append(item)
+
+    def __repr__(self):
+        return self.output.__repr__()
+
+    def __len__(self):
+        if self.iter_mode in [
+            SAM3Output.IterMode.ALL_STEPS_PER_STAGE,
+            SAM3Output.IterMode.LAST_STEP_PER_STAGE,
+        ]:
+            return len(self.output)
+        elif self.iter_mode == SAM3Output.IterMode.FLATTENED:
+            flattened_output = sum(self.output, [])
+            return len(flattened_output)

sam3/model/necks.py

@@ -0,0 +1,125 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+"""Necks are the interface between a vision backbone and the rest of the detection model"""
+
+from copy import deepcopy
+from typing import List, Optional, Tuple
+
+import torch
+
+import torch.nn as nn
+
+
+class Sam3DualViTDetNeck(nn.Module):
+    def __init__(
+        self,
+        trunk: nn.Module,
+        position_encoding: nn.Module,
+        d_model: int,
+        scale_factors=(4.0, 2.0, 1.0, 0.5),
+        add_sam2_neck: bool = False,
+    ):
+        """
+        SimpleFPN neck a la ViTDet
+        (From detectron2, very lightly adapted)
+        It supports a "dual neck" setting, where we have two identical necks (for SAM3 and SAM2), with different weights
+
+        :param trunk: the backbone
+        :param position_encoding: the positional encoding to use
+        :param d_model: the dimension of the model
+        """
+        super().__init__()
+        self.trunk = trunk
+        self.position_encoding = position_encoding
+        self.convs = nn.ModuleList()
+
+        self.scale_factors = scale_factors
+        use_bias = True
+        dim: int = self.trunk.channel_list[-1]
+
+        for _, scale in enumerate(scale_factors):
+            current = nn.Sequential()
+
+            if scale == 4.0:
+                current.add_module(
+                    "dconv_2x2_0",
+                    nn.ConvTranspose2d(dim, dim // 2, kernel_size=2, stride=2),
+                )
+                current.add_module(
+                    "gelu",
+                    nn.GELU(),
+                )
+                current.add_module(
+                    "dconv_2x2_1",
+                    nn.ConvTranspose2d(dim // 2, dim // 4, kernel_size=2, stride=2),
+                )
+                out_dim = dim // 4
+            elif scale == 2.0:
+                current.add_module(
+                    "dconv_2x2",
+                    nn.ConvTranspose2d(dim, dim // 2, kernel_size=2, stride=2),
+                )
+                out_dim = dim // 2
+            elif scale == 1.0:
+                out_dim = dim
+            elif scale == 0.5:
+                current.add_module(
+                    "maxpool_2x2",
+                    nn.MaxPool2d(kernel_size=2, stride=2),
+                )
+                out_dim = dim
+            else:
+                raise NotImplementedError(f"scale_factor={scale} is not supported yet.")
+
+            current.add_module(
+                "conv_1x1",
+                nn.Conv2d(
+                    in_channels=out_dim,
+                    out_channels=d_model,
+                    kernel_size=1,
+                    bias=use_bias,
+                ),
+            )
+            current.add_module(
+                "conv_3x3",
+                nn.Conv2d(
+                    in_channels=d_model,
+                    out_channels=d_model,
+                    kernel_size=3,
+                    padding=1,
+                    bias=use_bias,
+                ),
+            )
+            self.convs.append(current)
+
+        self.sam2_convs = None
+        if add_sam2_neck:
+            # Assumes sam2 neck is just a clone of the original neck
+            self.sam2_convs = deepcopy(self.convs)
+
+    def forward(
+        self, tensor_list: List[torch.Tensor]
+    ) -> Tuple[
+        List[torch.Tensor],
+        List[torch.Tensor],
+        Optional[List[torch.Tensor]],
+        Optional[List[torch.Tensor]],
+    ]:
+        xs = self.trunk(tensor_list)
+        sam3_out, sam3_pos = [], []
+        sam2_out, sam2_pos = None, None
+        if self.sam2_convs is not None:
+            sam2_out, sam2_pos = [], []
+        x = xs[-1]  # simpleFPN
+        for i in range(len(self.convs)):
+            sam3_x_out = self.convs[i](x)
+            sam3_pos_out = self.position_encoding(sam3_x_out).to(sam3_x_out.dtype)
+            sam3_out.append(sam3_x_out)
+            sam3_pos.append(sam3_pos_out)
+
+            if self.sam2_convs is not None:
+                sam2_x_out = self.sam2_convs[i](x)
+                sam2_pos_out = self.position_encoding(sam2_x_out).to(sam2_x_out.dtype)
+                sam2_out.append(sam2_x_out)
+                sam2_pos.append(sam2_pos_out)
+        return sam3_out, sam3_pos, sam2_out, sam2_pos

sam3/model/position_encoding.py

@@ -0,0 +1,124 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import math
+from typing import Optional
+
+import torch
+from torch import nn
+
+
+class PositionEmbeddingSine(nn.Module):
+    """
+    This is a more standard version of the position embedding, very similar to the one
+    used by the Attention is all you need paper, generalized to work on images.
+    """
+
+    def __init__(
+        self,
+        num_pos_feats,
+        temperature: int = 10000,
+        normalize: bool = True,
+        scale: Optional[float] = None,
+        precompute_resolution: Optional[int] = None,
+    ):
+        super().__init__()
+        assert num_pos_feats % 2 == 0, "Expecting even model width"
+        self.num_pos_feats = num_pos_feats // 2
+        self.temperature = temperature
+        self.normalize = normalize
+        if scale is not None and normalize is False:
+            raise ValueError("normalize should be True if scale is passed")
+        if scale is None:
+            scale = 2 * math.pi
+        self.scale = scale
+
+        self.cache = {}
+        # Precompute positional encodings under `precompute_resolution` to fill the cache
+        # and avoid symbolic shape tracing errors in torch.compile in PyTorch 2.4 nightly.
+        if precompute_resolution is not None:
+            # We precompute pos enc for stride 4, 8, 16 and 32 to fill `self.cache`.
+            precompute_sizes = [
+                (precompute_resolution // 4, precompute_resolution // 4),
+                (precompute_resolution // 8, precompute_resolution // 8),
+                (precompute_resolution // 16, precompute_resolution // 16),
+                (precompute_resolution // 32, precompute_resolution // 32),
+            ]
+            for size in precompute_sizes:
+                tensors = torch.zeros((1, 1) + size, device="cuda")
+                self.forward(tensors)
+                # further clone and detach it in the cache (just to be safe)
+                self.cache[size] = self.cache[size].clone().detach()
+
+    def _encode_xy(self, x, y):
+        # The positions are expected to be normalized
+        assert len(x) == len(y) and x.ndim == y.ndim == 1
+        x_embed = x * self.scale
+        y_embed = y * self.scale
+
+        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
+
+        pos_x = x_embed[:, None] / dim_t
+        pos_y = y_embed[:, None] / dim_t
+        pos_x = torch.stack(
+            (pos_x[:, 0::2].sin(), pos_x[:, 1::2].cos()), dim=2
+        ).flatten(1)
+        pos_y = torch.stack(
+            (pos_y[:, 0::2].sin(), pos_y[:, 1::2].cos()), dim=2
+        ).flatten(1)
+        return pos_x, pos_y
+
+    @torch.no_grad()
+    def encode_boxes(self, x, y, w, h):
+        pos_x, pos_y = self._encode_xy(x, y)
+        pos = torch.cat((pos_y, pos_x, h[:, None], w[:, None]), dim=1)
+        return pos
+
+    encode = encode_boxes  # Backwards compatibility
+
+    @torch.no_grad()
+    def encode_points(self, x, y, labels):
+        (bx, nx), (by, ny), (bl, nl) = x.shape, y.shape, labels.shape
+        assert bx == by and nx == ny and bx == bl and nx == nl
+        pos_x, pos_y = self._encode_xy(x.flatten(), y.flatten())
+        pos_x, pos_y = pos_x.reshape(bx, nx, -1), pos_y.reshape(by, ny, -1)
+        pos = torch.cat((pos_y, pos_x, labels[:, :, None]), dim=2)
+        return pos
+
+    @torch.no_grad()
+    def forward(self, x):
+        cache_key = None
+        cache_key = (x.shape[-2], x.shape[-1])
+        if cache_key in self.cache:
+            return self.cache[cache_key][None].repeat(x.shape[0], 1, 1, 1)
+        y_embed = (
+            torch.arange(1, x.shape[-2] + 1, dtype=torch.float32, device=x.device)
+            .view(1, -1, 1)
+            .repeat(x.shape[0], 1, x.shape[-1])
+        )
+        x_embed = (
+            torch.arange(1, x.shape[-1] + 1, dtype=torch.float32, device=x.device)
+            .view(1, 1, -1)
+            .repeat(x.shape[0], x.shape[-2], 1)
+        )
+
+        if self.normalize:
+            eps = 1e-6
+            y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
+            x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
+
+        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
+
+        pos_x = x_embed[:, :, :, None] / dim_t
+        pos_y = y_embed[:, :, :, None] / dim_t
+        pos_x = torch.stack(
+            (pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos_y = torch.stack(
+            (pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+        if cache_key is not None:
+            self.cache[cache_key] = pos[0]
+        return pos

sam3/model/sam1_task_predictor.py

@@ -0,0 +1,458 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+import torch.nn as nn
+from PIL.Image import Image
+
+from sam3.model.sam3_tracker_base import Sam3TrackerBase
+from sam3.model.utils.sam1_utils import SAM2Transforms
+
+
+# Adapted from https://github.com/facebookresearch/sam2/blob/main/sam2/sam2_image_predictor.py
+class SAM3InteractiveImagePredictor(nn.Module):
+    def __init__(
+        self,
+        sam_model: Sam3TrackerBase,
+        mask_threshold=0.0,
+        max_hole_area=256.0,
+        max_sprinkle_area=0.0,
+        **kwargs,
+    ) -> None:
+        """
+        Uses SAM-3 to calculate the image embedding for an image, and then
+        allow repeated, efficient mask prediction given prompts.
+
+        Arguments:
+          sam_model : The model to use for mask prediction.
+          mask_threshold (float): The threshold to use when converting mask logits
+            to binary masks. Masks are thresholded at 0 by default.
+          max_hole_area (int): If max_hole_area > 0, we fill small holes in up to
+            the maximum area of max_hole_area in low_res_masks.
+          max_sprinkle_area (int): If max_sprinkle_area > 0, we remove small sprinkles up to
+            the maximum area of max_sprinkle_area in low_res_masks.
+        """
+        super().__init__()
+        self.model = sam_model
+        self._transforms = SAM2Transforms(
+            resolution=self.model.image_size,
+            mask_threshold=mask_threshold,
+            max_hole_area=max_hole_area,
+            max_sprinkle_area=max_sprinkle_area,
+        )
+
+        # Predictor state
+        self._is_image_set = False
+        self._features = None
+        self._orig_hw = None
+        # Whether the predictor is set for single image or a batch of images
+        self._is_batch = False
+
+        # Predictor config
+        self.mask_threshold = mask_threshold
+
+        # Spatial dim for backbone feature maps
+        self._bb_feat_sizes = [
+            (288, 288),
+            (144, 144),
+            (72, 72),
+        ]
+
+    @torch.no_grad()
+    def set_image(
+        self,
+        image: Union[np.ndarray, Image],
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image, allowing
+        masks to be predicted with the 'predict' method.
+
+        Arguments:
+          image (np.ndarray or PIL Image): The input image to embed in RGB format. The image should be in HWC format if np.ndarray, or WHC format if PIL Image
+          with pixel values in [0, 255].
+          image_format (str): The color format of the image, in ['RGB', 'BGR'].
+        """
+        self.reset_predictor()
+        # Transform the image to the form expected by the model
+        if isinstance(image, np.ndarray):
+            logging.info("For numpy array image, we assume (HxWxC) format")
+            self._orig_hw = [image.shape[:2]]
+        elif isinstance(image, Image):
+            w, h = image.size
+            self._orig_hw = [(h, w)]
+        else:
+            raise NotImplementedError("Image format not supported")
+
+        input_image = self._transforms(image)
+        input_image = input_image[None, ...].to(self.device)
+
+        assert (
+            len(input_image.shape) == 4 and input_image.shape[1] == 3
+        ), f"input_image must be of size 1x3xHxW, got {input_image.shape}"
+        logging.info("Computing image embeddings for the provided image...")
+        backbone_out = self.model.forward_image(input_image)
+        (
+            _,
+            vision_feats,
+            _,
+            _,
+        ) = self.model._prepare_backbone_features(backbone_out)
+        # Add no_mem_embed, which is added to the lowest rest feat. map during training on videos
+        vision_feats[-1] = vision_feats[-1] + self.model.no_mem_embed
+
+        feats = [
+            feat.permute(1, 2, 0).view(1, -1, *feat_size)
+            for feat, feat_size in zip(vision_feats[::-1], self._bb_feat_sizes[::-1])
+        ][::-1]
+        self._features = {"image_embed": feats[-1], "high_res_feats": feats[:-1]}
+        self._is_image_set = True
+        logging.info("Image embeddings computed.")
+
+    @torch.no_grad()
+    def set_image_batch(
+        self,
+        image_list: List[Union[np.ndarray]],
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image batch, allowing
+        masks to be predicted with the 'predict_batch' method.
+
+        Arguments:
+          image_list (List[np.ndarray]): The input images to embed in RGB format. The image should be in HWC format if np.ndarray
+          with pixel values in [0, 255].
+        """
+        self.reset_predictor()
+        assert isinstance(image_list, list)
+        self._orig_hw = []
+        for image in image_list:
+            assert isinstance(
+                image, np.ndarray
+            ), "Images are expected to be an np.ndarray in RGB format, and of shape  HWC"
+            self._orig_hw.append(image.shape[:2])
+        # Transform the image to the form expected by the model
+        img_batch = self._transforms.forward_batch(image_list)
+        img_batch = img_batch.to(self.device)
+        batch_size = img_batch.shape[0]
+        assert (
+            len(img_batch.shape) == 4 and img_batch.shape[1] == 3
+        ), f"img_batch must be of size Bx3xHxW, got {img_batch.shape}"
+        logging.info("Computing image embeddings for the provided images...")
+        backbone_out = self.model.forward_image(img_batch)
+        (
+            _,
+            vision_feats,
+            _,
+            _,
+        ) = self.model._prepare_backbone_features(backbone_out)
+        # Add no_mem_embed, which is added to the lowest rest feat. map during training on videos
+        vision_feats[-1] = vision_feats[-1] + self.model.no_mem_embed
+
+        feats = [
+            feat.permute(1, 2, 0).view(batch_size, -1, *feat_size)
+            for feat, feat_size in zip(vision_feats[::-1], self._bb_feat_sizes[::-1])
+        ][::-1]
+        self._features = {"image_embed": feats[-1], "high_res_feats": feats[:-1]}
+        self._is_image_set = True
+        self._is_batch = True
+        logging.info("Image embeddings computed.")
+
+    def predict_batch(
+        self,
+        point_coords_batch: List[np.ndarray] = None,
+        point_labels_batch: List[np.ndarray] = None,
+        box_batch: List[np.ndarray] = None,
+        mask_input_batch: List[np.ndarray] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        normalize_coords=True,
+    ) -> Tuple[List[np.ndarray], List[np.ndarray], List[np.ndarray]]:
+        """This function is very similar to predict(...), however it is used for batched mode, when the model is expected to generate predictions on multiple images.
+        It returns a tuple of lists of masks, ious, and low_res_masks_logits.
+        """
+        assert self._is_batch, "This function should only be used when in batched mode"
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image_batch(...) before mask prediction."
+            )
+        num_images = len(self._features["image_embed"])
+        all_masks = []
+        all_ious = []
+        all_low_res_masks = []
+        for img_idx in range(num_images):
+            # Transform input prompts
+            point_coords = (
+                point_coords_batch[img_idx] if point_coords_batch is not None else None
+            )
+            point_labels = (
+                point_labels_batch[img_idx] if point_labels_batch is not None else None
+            )
+            box = box_batch[img_idx] if box_batch is not None else None
+            mask_input = (
+                mask_input_batch[img_idx] if mask_input_batch is not None else None
+            )
+            mask_input, unnorm_coords, labels, unnorm_box = self._prep_prompts(
+                point_coords,
+                point_labels,
+                box,
+                mask_input,
+                normalize_coords,
+                img_idx=img_idx,
+            )
+            masks, iou_predictions, low_res_masks = self._predict(
+                unnorm_coords,
+                labels,
+                unnorm_box,
+                mask_input,
+                multimask_output,
+                return_logits=return_logits,
+                img_idx=img_idx,
+            )
+            masks_np = masks.squeeze(0).float().detach().cpu().numpy()
+            iou_predictions_np = (
+                iou_predictions.squeeze(0).float().detach().cpu().numpy()
+            )
+            low_res_masks_np = low_res_masks.squeeze(0).float().detach().cpu().numpy()
+            all_masks.append(masks_np)
+            all_ious.append(iou_predictions_np)
+            all_low_res_masks.append(low_res_masks_np)
+
+        return all_masks, all_ious, all_low_res_masks
+
+    def predict(
+        self,
+        point_coords: Optional[np.ndarray] = None,
+        point_labels: Optional[np.ndarray] = None,
+        box: Optional[np.ndarray] = None,
+        mask_input: Optional[np.ndarray] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        normalize_coords=True,
+    ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+
+        Arguments:
+          point_coords (np.ndarray or None): A Nx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (np.ndarray or None): A length N array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          box (np.ndarray or None): A length 4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form 1xHxW, where
+            for SAM, H=W=256.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+          normalize_coords (bool): If true, the point coordinates will be normalized to the range [0,1] and point_coords is expected to be wrt. image dimensions.
+
+        Returns:
+          (np.ndarray): The output masks in CxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (np.ndarray): An array of length C containing the model's
+            predictions for the quality of each mask.
+          (np.ndarray): An array of shape CxHxW, where C is the number
+            of masks and H=W=256. These low resolution logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) before mask prediction."
+            )
+
+        # Transform input prompts
+
+        mask_input, unnorm_coords, labels, unnorm_box = self._prep_prompts(
+            point_coords, point_labels, box, mask_input, normalize_coords
+        )
+
+        masks, iou_predictions, low_res_masks = self._predict(
+            unnorm_coords,
+            labels,
+            unnorm_box,
+            mask_input,
+            multimask_output,
+            return_logits=return_logits,
+        )
+
+        masks_np = masks.squeeze(0).float().detach().cpu().numpy()
+        iou_predictions_np = iou_predictions.squeeze(0).float().detach().cpu().numpy()
+        low_res_masks_np = low_res_masks.squeeze(0).float().detach().cpu().numpy()
+        return masks_np, iou_predictions_np, low_res_masks_np
+
+    def _prep_prompts(
+        self, point_coords, point_labels, box, mask_logits, normalize_coords, img_idx=-1
+    ):
+        unnorm_coords, labels, unnorm_box, mask_input = None, None, None, None
+        if point_coords is not None:
+            assert (
+                point_labels is not None
+            ), "point_labels must be supplied if point_coords is supplied."
+            point_coords = torch.as_tensor(
+                point_coords, dtype=torch.float, device=self.device
+            )
+            unnorm_coords = self._transforms.transform_coords(
+                point_coords, normalize=normalize_coords, orig_hw=self._orig_hw[img_idx]
+            )
+            labels = torch.as_tensor(point_labels, dtype=torch.int, device=self.device)
+            if len(unnorm_coords.shape) == 2:
+                unnorm_coords, labels = unnorm_coords[None, ...], labels[None, ...]
+        if box is not None:
+            box = torch.as_tensor(box, dtype=torch.float, device=self.device)
+            unnorm_box = self._transforms.transform_boxes(
+                box, normalize=normalize_coords, orig_hw=self._orig_hw[img_idx]
+            )  # Bx2x2
+        if mask_logits is not None:
+            mask_input = torch.as_tensor(
+                mask_logits, dtype=torch.float, device=self.device
+            )
+            if len(mask_input.shape) == 3:
+                mask_input = mask_input[None, :, :, :]
+        return mask_input, unnorm_coords, labels, unnorm_box
+
+    @torch.no_grad()
+    def _predict(
+        self,
+        point_coords: Optional[torch.Tensor],
+        point_labels: Optional[torch.Tensor],
+        boxes: Optional[torch.Tensor] = None,
+        mask_input: Optional[torch.Tensor] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        img_idx: int = -1,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+        Input prompts are batched torch tensors and are expected to already be
+        transformed to the input frame using SAM2Transforms.
+
+        Arguments:
+          point_coords (torch.Tensor or None): A BxNx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (torch.Tensor or None): A BxN array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          boxes (np.ndarray or None): A Bx4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form Bx1xHxW, where
+            for SAM, H=W=256. Masks returned by a previous iteration of the
+            predict method do not need further transformation.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+
+        Returns:
+          (torch.Tensor): The output masks in BxCxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (torch.Tensor): An array of shape BxC containing the model's
+            predictions for the quality of each mask.
+          (torch.Tensor): An array of shape BxCxHxW, where C is the number
+            of masks and H=W=256. These low res logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) before mask prediction."
+            )
+
+        if point_coords is not None:
+            concat_points = (point_coords, point_labels)
+        else:
+            concat_points = None
+
+        # Embed prompts
+        if boxes is not None:
+            box_coords = boxes.reshape(-1, 2, 2)
+            box_labels = torch.tensor([[2, 3]], dtype=torch.int, device=boxes.device)
+            box_labels = box_labels.repeat(boxes.size(0), 1)
+            # we merge "boxes" and "points" into a single "concat_points" input (where
+            # boxes are added at the beginning) to sam_prompt_encoder
+            if concat_points is not None:
+                concat_coords = torch.cat([box_coords, concat_points[0]], dim=1)
+                concat_labels = torch.cat([box_labels, concat_points[1]], dim=1)
+                concat_points = (concat_coords, concat_labels)
+            else:
+                concat_points = (box_coords, box_labels)
+
+        sparse_embeddings, dense_embeddings = self.model.sam_prompt_encoder(
+            points=concat_points,
+            boxes=None,
+            masks=mask_input,
+        )
+
+        # Predict masks
+        batched_mode = (
+            concat_points is not None and concat_points[0].shape[0] > 1
+        )  # multi object prediction
+        high_res_features = [
+            feat_level[img_idx].unsqueeze(0)
+            for feat_level in self._features["high_res_feats"]
+        ]
+        low_res_masks, iou_predictions, _, _ = self.model.sam_mask_decoder(
+            image_embeddings=self._features["image_embed"][img_idx].unsqueeze(0),
+            image_pe=self.model.sam_prompt_encoder.get_dense_pe(),
+            sparse_prompt_embeddings=sparse_embeddings,
+            dense_prompt_embeddings=dense_embeddings,
+            multimask_output=multimask_output,
+            repeat_image=batched_mode,
+            high_res_features=high_res_features,
+        )
+
+        # Upscale the masks to the original image resolution
+        masks = self._transforms.postprocess_masks(
+            low_res_masks, self._orig_hw[img_idx]
+        )
+        low_res_masks = torch.clamp(low_res_masks, -32.0, 32.0)
+        if not return_logits:
+            masks = masks > self.mask_threshold
+
+        return masks, iou_predictions, low_res_masks
+
+    def get_image_embedding(self) -> torch.Tensor:
+        """
+        Returns the image embeddings for the currently set image, with
+        shape 1xCxHxW, where C is the embedding dimension and (H,W) are
+        the embedding spatial dimension of SAM (typically C=256, H=W=64).
+        """
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) to generate an embedding."
+            )
+        assert (
+            self._features is not None
+        ), "Features must exist if an image has been set."
+        return self._features["image_embed"]
+
+    @property
+    def device(self) -> torch.device:
+        return self.model.device
+
+    def reset_predictor(self) -> None:
+        """
+        Resets the image embeddings and other state variables.
+        """
+        self._is_image_set = False
+        self._features = None
+        self._orig_hw = None
+        self._is_batch = False

sam3/model/sam3_image.py

@@ -0,0 +1,883 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import os
+from copy import deepcopy
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+import torch
+
+from sam3.model.model_misc import SAM3Output
+
+from sam3.model.sam1_task_predictor import SAM3InteractiveImagePredictor
+from sam3.model.vl_combiner import SAM3VLBackbone
+from sam3.perflib.nms import nms_masks
+
+from sam3.train.data.collator import BatchedDatapoint
+
+from .act_ckpt_utils import activation_ckpt_wrapper
+
+from .box_ops import box_cxcywh_to_xyxy
+
+from .geometry_encoders import Prompt
+from .model_misc import inverse_sigmoid
+
+
+def _update_out(out, out_name, out_value, auxiliary=True, update_aux=True):
+    out[out_name] = out_value[-1] if auxiliary else out_value
+    if auxiliary and update_aux:
+        if "aux_outputs" not in out:
+            out["aux_outputs"] = [{} for _ in range(len(out_value) - 1)]
+        assert len(out["aux_outputs"]) == len(out_value) - 1
+        for aux_output, aux_value in zip(out["aux_outputs"], out_value[:-1]):
+            aux_output[out_name] = aux_value
+
+
+class Sam3Image(torch.nn.Module):
+    TEXT_ID_FOR_TEXT = 0
+    TEXT_ID_FOR_VISUAL = 1
+    TEXT_ID_FOR_GEOMETRIC = 2
+
+    def __init__(
+        self,
+        backbone: SAM3VLBackbone,
+        transformer,
+        input_geometry_encoder,
+        segmentation_head=None,
+        num_feature_levels=1,
+        o2m_mask_predict=True,
+        dot_prod_scoring=None,
+        use_instance_query: bool = True,
+        multimask_output: bool = True,
+        use_act_checkpoint_seg_head: bool = True,
+        interactivity_in_encoder: bool = True,
+        matcher=None,
+        use_dot_prod_scoring=True,
+        supervise_joint_box_scores: bool = False,  # only relevant if using presence token/score
+        detach_presence_in_joint_score: bool = False,  # only relevant if using presence token/score
+        separate_scorer_for_instance: bool = False,
+        num_interactive_steps_val: int = 0,
+        inst_interactive_predictor: SAM3InteractiveImagePredictor = None,
+        **kwargs,
+    ):
+        super().__init__()
+        self.backbone = backbone
+        self.geometry_encoder = input_geometry_encoder
+        self.transformer = transformer
+        self.hidden_dim = transformer.d_model
+        self.num_feature_levels = num_feature_levels
+        self.segmentation_head = segmentation_head
+
+        self.o2m_mask_predict = o2m_mask_predict
+
+        self.dot_prod_scoring = dot_prod_scoring
+        self.use_act_checkpoint_seg_head = use_act_checkpoint_seg_head
+        self.interactivity_in_encoder = interactivity_in_encoder
+        self.matcher = matcher
+
+        self.num_interactive_steps_val = num_interactive_steps_val
+        self.use_dot_prod_scoring = use_dot_prod_scoring
+
+        if self.use_dot_prod_scoring:
+            assert dot_prod_scoring is not None
+            self.dot_prod_scoring = dot_prod_scoring
+            self.instance_dot_prod_scoring = None
+            if separate_scorer_for_instance:
+                self.instance_dot_prod_scoring = deepcopy(dot_prod_scoring)
+        else:
+            self.class_embed = torch.nn.Linear(self.hidden_dim, 1)
+            self.instance_class_embed = None
+            if separate_scorer_for_instance:
+                self.instance_class_embed = deepcopy(self.class_embed)
+
+        self.supervise_joint_box_scores = supervise_joint_box_scores
+        self.detach_presence_in_joint_score = detach_presence_in_joint_score
+
+        # verify the number of queries for O2O and O2M
+        num_o2o_static = self.transformer.decoder.num_queries
+        num_o2m_static = self.transformer.decoder.num_o2m_queries
+        assert num_o2m_static == (num_o2o_static if self.transformer.decoder.dac else 0)
+        self.dac = self.transformer.decoder.dac
+
+        self.use_instance_query = use_instance_query
+        self.multimask_output = multimask_output
+
+        self.inst_interactive_predictor = inst_interactive_predictor
+
+    @property
+    def device(self):
+        self._device = getattr(self, "_device", None) or next(self.parameters()).device
+        return self._device
+
+    def to(self, *args, **kwargs):
+        # clear cached _device in case the model is moved to a different device
+        self._device = None
+        return super().to(*args, **kwargs)
+
+    def _get_img_feats(self, backbone_out, img_ids):
+        """Retrieve correct image features from backbone output."""
+        if "backbone_fpn" in backbone_out:
+            if "id_mapping" in backbone_out and backbone_out["id_mapping"] is not None:
+                img_ids = backbone_out["id_mapping"][img_ids]
+                # If this assert fails, it likely means we're requesting different img_ids (perhaps a different frame?)
+                # We currently don't expect this to happen. We could technically trigger a recompute here,
+                # but likely at the cost of a cpu<->gpu sync point, which would deteriorate perf
+                torch._assert_async((img_ids >= 0).all())
+
+            vis_feats = backbone_out["backbone_fpn"][-self.num_feature_levels :]
+            vis_pos_enc = backbone_out["vision_pos_enc"][-self.num_feature_levels :]
+            vis_feat_sizes = [x.shape[-2:] for x in vis_pos_enc]  # (H, W) shapes
+            # index and flatten visual features NxCxHxW => HWxNxC (batch-first => seq-first)
+            img_feats = [x[img_ids].flatten(2).permute(2, 0, 1) for x in vis_feats]
+            img_pos_embeds = [
+                x[img_ids].flatten(2).permute(2, 0, 1) for x in vis_pos_enc
+            ]
+            return backbone_out, img_feats, img_pos_embeds, vis_feat_sizes
+
+        # Image features not available in backbone output, so we compute them on the fly
+        # This case likely occurs for video. In that case, we want to forward only the current frame
+        img_batch = backbone_out["img_batch_all_stages"]
+        if img_ids.numel() > 1:
+            # Only forward backbone on unique image ids to avoid repetitive computation
+            unique_ids, _ = torch.unique(img_ids, return_inverse=True)
+        else:
+            unique_ids, _ = img_ids, slice(None)
+        # Compute the image features on those unique image ids
+        # note: we allow using a list (or other indexable types) of tensors as img_batch
+        # (e.g. for async frame loading in demo). In this case we index img_batch.tensors directly
+        if isinstance(img_batch, torch.Tensor):
+            image = img_batch[unique_ids]
+        elif unique_ids.numel() == 1:
+            image = img_batch[unique_ids.item()].unsqueeze(0)
+        else:
+            image = torch.stack([img_batch[i] for i in unique_ids.tolist()])
+        # `img_batch` might be fp16 and offloaded to CPU
+        image = image.to(dtype=torch.float32, device=self.device)
+        # Next time we call this function, we want to remember which indices we computed
+        id_mapping = torch.full(
+            (len(img_batch),), -1, dtype=torch.long, device=self.device
+        )
+        id_mapping[unique_ids] = torch.arange(len(unique_ids), device=self.device)
+        backbone_out = {
+            **backbone_out,
+            **self.backbone.forward_image(image),
+            "id_mapping": id_mapping,
+        }
+        assert "backbone_fpn" in backbone_out
+        return self._get_img_feats(backbone_out, img_ids=img_ids)
+
+    def _encode_prompt(
+        self,
+        backbone_out,
+        find_input,
+        geometric_prompt,
+        visual_prompt_embed=None,
+        visual_prompt_mask=None,
+        encode_text=True,
+        prev_mask_pred=None,
+    ):
+        # index text features (note that regardless of early or late fusion, the batch size of
+        # `txt_feats` is always the number of *prompts* in the encoder)
+        txt_ids = find_input.text_ids
+        txt_feats = backbone_out["language_features"][:, txt_ids]
+        txt_masks = backbone_out["language_mask"][txt_ids]
+
+        feat_tuple = self._get_img_feats(backbone_out, find_input.img_ids)
+        backbone_out, img_feats, img_pos_embeds, vis_feat_sizes = feat_tuple
+
+        if prev_mask_pred is not None:
+            img_feats = [img_feats[-1] + prev_mask_pred]
+        # Encode geometry
+        geo_feats, geo_masks = self.geometry_encoder(
+            geo_prompt=geometric_prompt,
+            img_feats=img_feats,
+            img_sizes=vis_feat_sizes,
+            img_pos_embeds=img_pos_embeds,
+        )
+        if visual_prompt_embed is None:
+            visual_prompt_embed = torch.zeros(
+                (0, *geo_feats.shape[1:]), device=geo_feats.device
+            )
+            visual_prompt_mask = torch.zeros(
+                (*geo_masks.shape[:-1], 0),
+                device=geo_masks.device,
+                dtype=geo_masks.dtype,
+            )
+        if encode_text:
+            prompt = torch.cat([txt_feats, geo_feats, visual_prompt_embed], dim=0)
+            prompt_mask = torch.cat([txt_masks, geo_masks, visual_prompt_mask], dim=1)
+        else:
+            prompt = torch.cat([geo_feats, visual_prompt_embed], dim=0)
+            prompt_mask = torch.cat([geo_masks, visual_prompt_mask], dim=1)
+        return prompt, prompt_mask, backbone_out
+
+    def _run_encoder(
+        self,
+        backbone_out,
+        find_input,
+        prompt,
+        prompt_mask,
+        encoder_extra_kwargs: Optional[Dict] = None,
+    ):
+        feat_tuple = self._get_img_feats(backbone_out, find_input.img_ids)
+        backbone_out, img_feats, img_pos_embeds, vis_feat_sizes = feat_tuple
+
+        # Run the encoder
+        prompt_pos_embed = torch.zeros_like(prompt)
+        # make a copy of the image feature lists since the encoder may modify these lists in-place
+        memory = self.transformer.encoder(
+            src=img_feats.copy(),
+            src_key_padding_mask=None,
+            src_pos=img_pos_embeds.copy(),
+            prompt=prompt,
+            prompt_pos=prompt_pos_embed,
+            prompt_key_padding_mask=prompt_mask,
+            feat_sizes=vis_feat_sizes,
+            encoder_extra_kwargs=encoder_extra_kwargs,
+        )
+        encoder_out = {
+            # encoded image features
+            "encoder_hidden_states": memory["memory"],
+            "pos_embed": memory["pos_embed"],
+            "padding_mask": memory["padding_mask"],
+            "level_start_index": memory["level_start_index"],
+            "spatial_shapes": memory["spatial_shapes"],
+            "valid_ratios": memory["valid_ratios"],
+            "vis_feat_sizes": vis_feat_sizes,
+            # encoded text features (or other prompts)
+            "prompt_before_enc": prompt,
+            "prompt_after_enc": memory.get("memory_text", prompt),
+            "prompt_mask": prompt_mask,
+        }
+        return backbone_out, encoder_out, feat_tuple
+
+    def _run_decoder(
+        self,
+        pos_embed,
+        memory,
+        src_mask,
+        out,
+        prompt,
+        prompt_mask,
+        encoder_out,
+    ):
+        bs = memory.shape[1]
+        query_embed = self.transformer.decoder.query_embed.weight
+        tgt = query_embed.unsqueeze(1).repeat(1, bs, 1)
+
+        apply_dac = self.transformer.decoder.dac and self.training
+        hs, reference_boxes, dec_presence_out, dec_presence_feats = (
+            self.transformer.decoder(
+                tgt=tgt,
+                memory=memory,
+                memory_key_padding_mask=src_mask,
+                pos=pos_embed,
+                reference_boxes=None,
+                level_start_index=encoder_out["level_start_index"],
+                spatial_shapes=encoder_out["spatial_shapes"],
+                valid_ratios=encoder_out["valid_ratios"],
+                tgt_mask=None,
+                memory_text=prompt,
+                text_attention_mask=prompt_mask,
+                apply_dac=apply_dac,
+            )
+        )
+        hs = hs.transpose(1, 2)  # seq-first to batch-first
+        reference_boxes = reference_boxes.transpose(1, 2)  # seq-first to batch-first
+        if dec_presence_out is not None:
+            # seq-first to batch-first
+            dec_presence_out = dec_presence_out.transpose(1, 2)
+
+        out["presence_feats"] = dec_presence_feats
+        self._update_scores_and_boxes(
+            out,
+            hs,
+            reference_boxes,
+            prompt,
+            prompt_mask,
+            dec_presence_out=dec_presence_out,
+        )
+        return out, hs
+
+    def _update_scores_and_boxes(
+        self,
+        out,
+        hs,
+        reference_boxes,
+        prompt,
+        prompt_mask,
+        dec_presence_out=None,
+        is_instance_prompt=False,
+    ):
+        apply_dac = self.transformer.decoder.dac and self.training
+        num_o2o = (hs.size(2) // 2) if apply_dac else hs.size(2)
+        num_o2m = hs.size(2) - num_o2o
+        assert num_o2m == (num_o2o if apply_dac else 0)
+        out["queries"] = hs[-1][:, :num_o2o]  # remove o2m queries if there are any
+        # score prediction
+        if self.use_dot_prod_scoring:
+            dot_prod_scoring_head = self.dot_prod_scoring
+            if is_instance_prompt and self.instance_dot_prod_scoring is not None:
+                dot_prod_scoring_head = self.instance_dot_prod_scoring
+            outputs_class = dot_prod_scoring_head(hs, prompt, prompt_mask)
+        else:
+            class_embed_head = self.class_embed
+            if is_instance_prompt and self.instance_class_embed is not None:
+                class_embed_head = self.instance_class_embed
+            outputs_class = class_embed_head(hs)
+
+        # box prediction
+        box_head = self.transformer.decoder.bbox_embed
+        if (
+            is_instance_prompt
+            and self.transformer.decoder.instance_bbox_embed is not None
+        ):
+            box_head = self.transformer.decoder.instance_bbox_embed
+        anchor_box_offsets = box_head(hs)
+        reference_boxes_inv_sig = inverse_sigmoid(reference_boxes)
+        outputs_coord = (reference_boxes_inv_sig + anchor_box_offsets).sigmoid()
+        outputs_boxes_xyxy = box_cxcywh_to_xyxy(outputs_coord)
+
+        if dec_presence_out is not None:
+            _update_out(
+                out, "presence_logit_dec", dec_presence_out, update_aux=self.training
+            )
+
+        if self.supervise_joint_box_scores:
+            assert dec_presence_out is not None
+            prob_dec_presence_out = dec_presence_out.clone().sigmoid()
+            if self.detach_presence_in_joint_score:
+                prob_dec_presence_out = prob_dec_presence_out.detach()
+
+            outputs_class = inverse_sigmoid(
+                outputs_class.sigmoid() * prob_dec_presence_out.unsqueeze(2)
+            ).clamp(min=-10.0, max=10.0)
+
+        _update_out(
+            out, "pred_logits", outputs_class[:, :, :num_o2o], update_aux=self.training
+        )
+        _update_out(
+            out, "pred_boxes", outputs_coord[:, :, :num_o2o], update_aux=self.training
+        )
+        _update_out(
+            out,
+            "pred_boxes_xyxy",
+            outputs_boxes_xyxy[:, :, :num_o2o],
+            update_aux=self.training,
+        )
+        if num_o2m > 0 and self.training:
+            _update_out(
+                out,
+                "pred_logits_o2m",
+                outputs_class[:, :, num_o2o:],
+                update_aux=self.training,
+            )
+            _update_out(
+                out,
+                "pred_boxes_o2m",
+                outputs_coord[:, :, num_o2o:],
+                update_aux=self.training,
+            )
+            _update_out(
+                out,
+                "pred_boxes_xyxy_o2m",
+                outputs_boxes_xyxy[:, :, num_o2o:],
+                update_aux=self.training,
+            )
+
+    def _run_segmentation_heads(
+        self,
+        out,
+        backbone_out,
+        img_ids,
+        vis_feat_sizes,
+        encoder_hidden_states,
+        prompt,
+        prompt_mask,
+        hs,
+    ):
+        apply_dac = self.transformer.decoder.dac and self.training
+        if self.segmentation_head is not None:
+            num_o2o = (hs.size(2) // 2) if apply_dac else hs.size(2)
+            num_o2m = hs.size(2) - num_o2o
+            obj_queries = hs if self.o2m_mask_predict else hs[:, :, :num_o2o]
+            seg_head_outputs = activation_ckpt_wrapper(self.segmentation_head)(
+                backbone_feats=backbone_out["backbone_fpn"],
+                obj_queries=obj_queries,
+                image_ids=img_ids,
+                encoder_hidden_states=encoder_hidden_states,
+                act_ckpt_enable=self.training and self.use_act_checkpoint_seg_head,
+                prompt=prompt,
+                prompt_mask=prompt_mask,
+            )
+            aux_masks = False  # self.aux_loss and self.segmentation_head.aux_masks
+            for k, v in seg_head_outputs.items():
+                if k in self.segmentation_head.instance_keys:
+                    _update_out(out, k, v[:, :num_o2o], auxiliary=aux_masks)
+                    if (
+                        self.o2m_mask_predict and num_o2m > 0
+                    ):  # handle o2m mask prediction
+                        _update_out(
+                            out, f"{k}_o2m", v[:, num_o2o:], auxiliary=aux_masks
+                        )
+                else:
+                    out[k] = v
+        else:
+            backbone_out.pop("backbone_fpn", None)
+
+    def _get_best_mask(self, out):
+        prev_mask_idx = out["pred_logits"].argmax(dim=1).squeeze(1)
+        batch_idx = torch.arange(
+            out["pred_logits"].shape[0], device=prev_mask_idx.device
+        )
+        prev_mask_pred = out["pred_masks"][batch_idx, prev_mask_idx][:, None]
+        # Downsample mask to match image resolution.
+        prev_mask_pred = self.geometry_encoder.mask_encoder.mask_downsampler(
+            prev_mask_pred
+        )
+        prev_mask_pred = prev_mask_pred.flatten(-2).permute(2, 0, 1)
+
+        return prev_mask_pred
+
+    def forward_grounding(
+        self,
+        backbone_out,
+        find_input,
+        find_target,
+        geometric_prompt: Prompt,
+    ):
+        with torch.profiler.record_function("SAM3Image._encode_prompt"):
+            prompt, prompt_mask, backbone_out = self._encode_prompt(
+                backbone_out, find_input, geometric_prompt
+            )
+        # Run the encoder
+        with torch.profiler.record_function("SAM3Image._run_encoder"):
+            backbone_out, encoder_out, _ = self._run_encoder(
+                backbone_out, find_input, prompt, prompt_mask
+            )
+        out = {
+            "encoder_hidden_states": encoder_out["encoder_hidden_states"],
+            "prev_encoder_out": {
+                "encoder_out": encoder_out,
+                "backbone_out": backbone_out,
+            },
+        }
+
+        # Run the decoder
+        with torch.profiler.record_function("SAM3Image._run_decoder"):
+            out, hs = self._run_decoder(
+                memory=out["encoder_hidden_states"],
+                pos_embed=encoder_out["pos_embed"],
+                src_mask=encoder_out["padding_mask"],
+                out=out,
+                prompt=prompt,
+                prompt_mask=prompt_mask,
+                encoder_out=encoder_out,
+            )
+
+        # Run segmentation heads
+        with torch.profiler.record_function("SAM3Image._run_segmentation_heads"):
+            self._run_segmentation_heads(
+                out=out,
+                backbone_out=backbone_out,
+                img_ids=find_input.img_ids,
+                vis_feat_sizes=encoder_out["vis_feat_sizes"],
+                encoder_hidden_states=out["encoder_hidden_states"],
+                prompt=prompt,
+                prompt_mask=prompt_mask,
+                hs=hs,
+            )
+
+        if self.training or self.num_interactive_steps_val > 0:
+            self._compute_matching(out, self.back_convert(find_target))
+        return out
+
+    def _postprocess_out(self, out: Dict, multimask_output: bool = False):
+        # For multimask output, during eval we return the single best mask with the dict keys expected by the evaluators, but also return the multimasks output with new keys.
+        num_mask_boxes = out["pred_boxes"].size(1)
+        if not self.training and multimask_output and num_mask_boxes > 1:
+            out["multi_pred_logits"] = out["pred_logits"]
+            if "pred_masks" in out:
+                out["multi_pred_masks"] = out["pred_masks"]
+            out["multi_pred_boxes"] = out["pred_boxes"]
+            out["multi_pred_boxes_xyxy"] = out["pred_boxes_xyxy"]
+
+            best_mask_idx = out["pred_logits"].argmax(1).squeeze(1)
+            batch_idx = torch.arange(len(best_mask_idx), device=best_mask_idx.device)
+
+            out["pred_logits"] = out["pred_logits"][batch_idx, best_mask_idx].unsqueeze(
+                1
+            )
+            if "pred_masks" in out:
+                out["pred_masks"] = out["pred_masks"][
+                    batch_idx, best_mask_idx
+                ].unsqueeze(1)
+            out["pred_boxes"] = out["pred_boxes"][batch_idx, best_mask_idx].unsqueeze(1)
+            out["pred_boxes_xyxy"] = out["pred_boxes_xyxy"][
+                batch_idx, best_mask_idx
+            ].unsqueeze(1)
+
+        return out
+
+    def _get_dummy_prompt(self, num_prompts=1):
+        device = self.device
+        geometric_prompt = Prompt(
+            box_embeddings=torch.zeros(0, num_prompts, 4, device=device),
+            box_mask=torch.zeros(num_prompts, 0, device=device, dtype=torch.bool),
+        )
+        return geometric_prompt
+
+    def forward(self, input: BatchedDatapoint):
+        device = self.device
+        backbone_out = {"img_batch_all_stages": input.img_batch}
+        backbone_out.update(self.backbone.forward_image(input.img_batch))
+        num_frames = len(input.find_inputs)
+        assert num_frames == 1
+
+        text_outputs = self.backbone.forward_text(input.find_text_batch, device=device)
+        backbone_out.update(text_outputs)
+
+        previous_stages_out = SAM3Output(
+            iter_mode=SAM3Output.IterMode.LAST_STEP_PER_STAGE
+        )
+
+        find_input = input.find_inputs[0]
+        find_target = input.find_targets[0]
+
+        if find_input.input_points is not None and find_input.input_points.numel() > 0:
+            print("Warning: Point prompts are ignored in PCS.")
+
+        num_interactive_steps = 0 if self.training else self.num_interactive_steps_val
+        geometric_prompt = Prompt(
+            box_embeddings=find_input.input_boxes,
+            box_mask=find_input.input_boxes_mask,
+            box_labels=find_input.input_boxes_label,
+        )
+
+        # Init vars that are shared across the loop.
+        stage_outs = []
+        for cur_step in range(num_interactive_steps + 1):
+            if cur_step > 0:
+                # We sample interactive geometric prompts (boxes, points)
+                geometric_prompt, _ = self.interactive_prompt_sampler.sample(
+                    geo_prompt=geometric_prompt,
+                    find_target=find_target,
+                    previous_out=stage_outs[-1],
+                )
+            out = self.forward_grounding(
+                backbone_out=backbone_out,
+                find_input=find_input,
+                find_target=find_target,
+                geometric_prompt=geometric_prompt.clone(),
+            )
+            stage_outs.append(out)
+
+        previous_stages_out.append(stage_outs)
+        return previous_stages_out
+
+    def _compute_matching(self, out, targets):
+        out["indices"] = self.matcher(out, targets)
+        for aux_out in out.get("aux_outputs", []):
+            aux_out["indices"] = self.matcher(aux_out, targets)
+
+    def back_convert(self, targets):
+        batched_targets = {
+            "boxes": targets.boxes.view(-1, 4),
+            "boxes_xyxy": box_cxcywh_to_xyxy(targets.boxes.view(-1, 4)),
+            "boxes_padded": targets.boxes_padded,
+            "positive_map": targets.boxes.new_ones(len(targets.boxes), 1),
+            "num_boxes": targets.num_boxes,
+            "masks": targets.segments,
+            "semantic_masks": targets.semantic_segments,
+            "is_valid_mask": targets.is_valid_segment,
+            "is_exhaustive": targets.is_exhaustive,
+            "object_ids_packed": targets.object_ids,
+            "object_ids_padded": targets.object_ids_padded,
+        }
+        return batched_targets
+
+    def predict_inst(
+        self,
+        inference_state,
+        **kwargs,
+    ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        orig_h, orig_w = (
+            inference_state["original_height"],
+            inference_state["original_width"],
+        )
+        backbone_out = inference_state["backbone_out"]["sam2_backbone_out"]
+        (
+            _,
+            vision_feats,
+            _,
+            _,
+        ) = self.inst_interactive_predictor.model._prepare_backbone_features(
+            backbone_out
+        )
+        # Add no_mem_embed, which is added to the lowest rest feat. map during training on videos
+        vision_feats[-1] = (
+            vision_feats[-1] + self.inst_interactive_predictor.model.no_mem_embed
+        )
+        feats = [
+            feat.permute(1, 2, 0).view(1, -1, *feat_size)
+            for feat, feat_size in zip(
+                vision_feats[::-1], self.inst_interactive_predictor._bb_feat_sizes[::-1]
+            )
+        ][::-1]
+        self.inst_interactive_predictor._features = {
+            "image_embed": feats[-1],
+            "high_res_feats": feats[:-1],
+        }
+        self.inst_interactive_predictor._is_image_set = True
+        self.inst_interactive_predictor._orig_hw = [(orig_h, orig_w)]
+        res = self.inst_interactive_predictor.predict(**kwargs)
+        self.inst_interactive_predictor._features = None
+        self.inst_interactive_predictor._is_image_set = False
+        return res
+
+    def predict_inst_batch(
+        self,
+        inference_state,
+        *args,
+        **kwargs,
+    ) -> Tuple[List[np.ndarray], List[np.ndarray], List[np.ndarray]]:
+        backbone_out = inference_state["backbone_out"]["sam2_backbone_out"]
+        (
+            _,
+            vision_feats,
+            _,
+            _,
+        ) = self.inst_interactive_predictor.model._prepare_backbone_features(
+            backbone_out
+        )
+        # Add no_mem_embed, which is added to the lowest res feat. map during training on videos
+        vision_feats[-1] = (
+            vision_feats[-1] + self.inst_interactive_predictor.model.no_mem_embed
+        )
+        batch_size = vision_feats[-1].shape[1]
+        orig_heights, orig_widths = (
+            inference_state["original_heights"],
+            inference_state["original_widths"],
+        )
+        assert (
+            batch_size == len(orig_heights) == len(orig_widths)
+        ), f"Batch size mismatch in predict_inst_batch. Got {batch_size}, {len(orig_heights)}, {len(orig_widths)}"
+        feats = [
+            feat.permute(1, 2, 0).view(batch_size, -1, *feat_size)
+            for feat, feat_size in zip(
+                vision_feats[::-1], self.inst_interactive_predictor._bb_feat_sizes[::-1]
+            )
+        ][::-1]
+        self.inst_interactive_predictor._features = {
+            "image_embed": feats[-1],
+            "high_res_feats": feats[:-1],
+        }
+        self.inst_interactive_predictor._is_image_set = True
+        self.inst_interactive_predictor._is_batch = True
+        self.inst_interactive_predictor._orig_hw = [
+            (orig_h, orig_w) for orig_h, orig_w in zip(orig_heights, orig_widths)
+        ]
+        res = self.inst_interactive_predictor.predict_batch(*args, **kwargs)
+        self.inst_interactive_predictor._features = None
+        self.inst_interactive_predictor._is_image_set = False
+        self.inst_interactive_predictor._is_batch = False
+        return res
+
+
+class Sam3ImageOnVideoMultiGPU(Sam3Image):
+    def __init__(
+        self, *args, async_all_gather=True, gather_backbone_out=None, **kwargs
+    ):
+        super().__init__(*args, **kwargs)
+        self.rank = int(os.getenv("RANK", "0"))
+        self.world_size = int(os.getenv("WORLD_SIZE", "1"))
+        self.async_all_gather = async_all_gather
+
+        # if gather_backbone is not set, default to gathering only for `SAM3VLBackbone`
+        if gather_backbone_out is None:
+            gather_backbone_out = isinstance(self.backbone, SAM3VLBackbone)
+        self.gather_backbone_out = gather_backbone_out
+
+    def forward_video_grounding_multigpu(
+        self,
+        backbone_out,
+        find_inputs,
+        geometric_prompt: Prompt,
+        frame_idx,
+        num_frames,
+        # `multigpu_buffer` is a dict to cache detector's outputs in a chunk between different calls
+        multigpu_buffer,
+        track_in_reverse=False,
+        # whether to also return the SAM2 backbone features
+        return_sam2_backbone_feats=False,
+        # whether to perform NMS and suppress the scores of those detections removed by NMS
+        run_nms=False,
+        nms_prob_thresh=None,
+        nms_iou_thresh=None,
+        **kwargs,
+    ):
+        """
+        Compute the detector's detection outputs in a distributed manner, where all GPUs process
+        a chunk of frames (equal to the number of GPUs) at once and store them in cache.
+        """
+        # Step 1: fetch the detector outputs in the current chunk from buffer
+        frame_idx_curr_b = frame_idx - frame_idx % self.world_size
+        frame_idx_curr_e = min(frame_idx_curr_b + self.world_size, num_frames)
+        # in case the current frame's detection results are not in the buffer yet, build the current chunk
+        # (this should only happen on the first chunk, since we are also building the next chunk below)
+        if frame_idx not in multigpu_buffer:
+            with torch.profiler.record_function("build_multigpu_buffer_next_chunk1"):
+                self._build_multigpu_buffer_next_chunk(
+                    backbone_out=backbone_out,
+                    find_inputs=find_inputs,
+                    geometric_prompt=geometric_prompt,
+                    frame_idx_begin=frame_idx_curr_b,
+                    frame_idx_end=frame_idx_curr_e,
+                    num_frames=num_frames,
+                    multigpu_buffer=multigpu_buffer,
+                    run_nms=run_nms,
+                    nms_prob_thresh=nms_prob_thresh,
+                    nms_iou_thresh=nms_iou_thresh,
+                )
+
+        # read out the current frame's results from `multigpu_buffer`
+        out = {}
+        for k, (v, handle) in multigpu_buffer[frame_idx].items():
+            if k.startswith("sam2_backbone_") and not return_sam2_backbone_feats:
+                continue
+            if handle is not None:
+                handle.wait()  # wait for async all-gather to finish
+            out[k] = v
+
+        # Step 2: remove detection outputs of the previous chunk from cache to save GPU memory
+        if not track_in_reverse and frame_idx_curr_b - self.world_size >= 0:
+            frame_idx_prev_e = frame_idx_curr_b
+            frame_idx_prev_b = frame_idx_curr_b - self.world_size
+        elif track_in_reverse and frame_idx_curr_e < num_frames:
+            frame_idx_prev_b = frame_idx_curr_e
+            frame_idx_prev_e = min(frame_idx_prev_b + self.world_size, num_frames)
+        else:
+            frame_idx_prev_b = frame_idx_prev_e = None
+        if frame_idx_prev_b is not None:
+            for frame_idx_rm in range(frame_idx_prev_b, frame_idx_prev_e):
+                multigpu_buffer.pop(frame_idx_rm, None)
+
+        # Step 3: compute and cache detection outputs of the next chunk ahead of time
+        # (so that we can overlap computation with all-gather transfer)
+        if not track_in_reverse and frame_idx_curr_e < num_frames:
+            frame_idx_next_b = frame_idx_curr_e
+            frame_idx_next_e = min(frame_idx_next_b + self.world_size, num_frames)
+        elif track_in_reverse and frame_idx_curr_b - self.world_size >= 0:
+            frame_idx_next_e = frame_idx_curr_b
+            frame_idx_next_b = frame_idx_curr_b - self.world_size
+        else:
+            frame_idx_next_b = frame_idx_next_e = None
+        if frame_idx_next_b is not None and frame_idx_next_b not in multigpu_buffer:
+            with torch.profiler.record_function("build_multigpu_buffer_next_chunk2"):
+                self._build_multigpu_buffer_next_chunk(
+                    backbone_out=backbone_out,
+                    find_inputs=find_inputs,
+                    geometric_prompt=geometric_prompt,
+                    frame_idx_begin=frame_idx_next_b,
+                    frame_idx_end=frame_idx_next_e,
+                    num_frames=num_frames,
+                    multigpu_buffer=multigpu_buffer,
+                    run_nms=run_nms,
+                    nms_prob_thresh=nms_prob_thresh,
+                    nms_iou_thresh=nms_iou_thresh,
+                )
+
+        return out, backbone_out
+
+    def _build_multigpu_buffer_next_chunk(
+        self,
+        backbone_out,
+        find_inputs,
+        geometric_prompt: Prompt,
+        frame_idx_begin,
+        frame_idx_end,
+        num_frames,
+        multigpu_buffer,
+        run_nms=False,
+        nms_prob_thresh=None,
+        nms_iou_thresh=None,
+    ):
+        """Compute detection outputs on a chunk of frames and store their results in multigpu_buffer."""
+        # each GPU computes detections on one frame in the chunk (in a round-robin manner)
+        frame_idx_local_gpu = min(frame_idx_begin + self.rank, frame_idx_end - 1)
+        # `forward_grounding` (from base class `Sam3ImageOnVideo`) runs the detector on a single frame
+        with torch.profiler.record_function("forward_grounding"):
+            out_local = self.forward_grounding(
+                backbone_out=backbone_out,
+                find_input=find_inputs[frame_idx_local_gpu],
+                find_target=None,
+                geometric_prompt=geometric_prompt,
+            )
+        if run_nms:
+            with torch.profiler.record_function("nms_masks"):
+                # run NMS as a post-processing step on top of the detection outputs
+                assert nms_prob_thresh is not None and nms_iou_thresh is not None
+                pred_probs = out_local["pred_logits"].squeeze(-1).sigmoid()
+                pred_masks = out_local["pred_masks"]
+                # loop over text prompts (not an overhead for demo where there's only 1 prompt)
+                for prompt_idx in range(pred_probs.size(0)):
+                    keep = nms_masks(
+                        pred_probs=pred_probs[prompt_idx],
+                        pred_masks=pred_masks[prompt_idx],
+                        prob_threshold=nms_prob_thresh,
+                        iou_threshold=nms_iou_thresh,
+                    )
+                    # set a very low threshold for those detections removed by NMS
+                    out_local["pred_logits"][prompt_idx, :, 0] -= 1e4 * (~keep).float()
+
+        if self.gather_backbone_out:
+            # gather the SAM 2 backbone features across GPUs
+            feats = out_local["prev_encoder_out"]["backbone_out"]["sam2_backbone_out"]
+            assert len(feats["backbone_fpn"]) == 3  # SAM2 backbone always have 3 levels
+            # cast the SAM2 backbone features to bfloat16 for all-gather (this is usually
+            # a no-op, SAM2 backbone features are likely already in bfloat16 due to AMP)
+            backbone_fpn_bf16 = [x.to(torch.bfloat16) for x in feats["backbone_fpn"]]
+            fpn0, fpn_handle0 = self._gather_tensor(backbone_fpn_bf16[0])
+            fpn1, fpn_handle1 = self._gather_tensor(backbone_fpn_bf16[1])
+            fpn2, fpn_handle2 = self._gather_tensor(backbone_fpn_bf16[2])
+            # vision_pos_enc is the same on all frames, so no need to all-gather them
+            vision_pos_enc = feats["vision_pos_enc"]
+
+        # trim the detector output to only include the necessary keys
+        out_local = {
+            "pred_logits": out_local["pred_logits"],
+            "pred_boxes": out_local["pred_boxes"],
+            "pred_boxes_xyxy": out_local["pred_boxes_xyxy"],
+            "pred_masks": out_local["pred_masks"],
+        }
+
+        # gather the results: after this step, each GPU will receive detector outputs on
+        # all frames in the chunk and store them in `multigpu_buffer`
+        out_gathered = {k: self._gather_tensor(v) for k, v in out_local.items()}
+        for rank in range(self.world_size):
+            frame_idx_to_save = frame_idx_begin + rank
+            if frame_idx_to_save >= num_frames:
+                continue
+            frame_buffer = {
+                k: (v[rank], handle) for k, (v, handle) in out_gathered.items()
+            }
+            if self.gather_backbone_out:
+                # also add gathered SAM 2 backbone features to frame_buffer
+                frame_buffer["tracker_backbone_fpn_0"] = (fpn0[rank], fpn_handle0)
+                frame_buffer["tracker_backbone_fpn_1"] = (fpn1[rank], fpn_handle1)
+                frame_buffer["tracker_backbone_fpn_2"] = (fpn2[rank], fpn_handle2)
+                frame_buffer["tracker_backbone_pos_enc"] = (vision_pos_enc, None)
+
+            multigpu_buffer[frame_idx_to_save] = frame_buffer
+
+    def _gather_tensor(self, x):
+        if self.world_size == 1:
+            return [x], None
+
+        async_op = self.async_all_gather
+        # here `.contiguous()` is required -- otherwise NCCL all_gather
+        # sometimes gives wrong results
+        x = x.contiguous()  # ensure contiguous memory for NCCL
+        output_list = [torch.empty_like(x) for _ in range(self.world_size)]
+        handle = torch.distributed.all_gather(output_list, x, async_op=async_op)
+        return output_list, handle

sam3/model/sam3_image_processor.py

@@ -0,0 +1,222 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+from typing import Dict, List
+
+import numpy as np
+import PIL
+import torch
+
+from sam3.model import box_ops
+
+from sam3.model.data_misc import FindStage, interpolate
+from torchvision.transforms import v2
+
+
+class Sam3Processor:
+    """ """
+
+    def __init__(self, model, resolution=1008, device="cuda", confidence_threshold=0.5):
+        self.model = model
+        self.resolution = resolution
+        self.device = device
+        self.transform = v2.Compose(
+            [
+                v2.ToDtype(torch.uint8, scale=True),
+                v2.Resize(size=(resolution, resolution)),
+                v2.ToDtype(torch.float32, scale=True),
+                v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+            ]
+        )
+        self.confidence_threshold = confidence_threshold
+
+        self.find_stage = FindStage(
+            img_ids=torch.tensor([0], device=device, dtype=torch.long),
+            text_ids=torch.tensor([0], device=device, dtype=torch.long),
+            input_boxes=None,
+            input_boxes_mask=None,
+            input_boxes_label=None,
+            input_points=None,
+            input_points_mask=None,
+        )
+
+    @torch.inference_mode()
+    def set_image(self, image, state=None):
+        """Sets the image on which we want to do predictions."""
+        if state is None:
+            state = {}
+
+        if isinstance(image, PIL.Image.Image):
+            width, height = image.size
+        elif isinstance(image, (torch.Tensor, np.ndarray)):
+            height, width = image.shape[-2:]
+        else:
+            raise ValueError("Image must be a PIL image or a tensor")
+
+        image = v2.functional.to_image(image).to(self.device)
+        image = self.transform(image).unsqueeze(0)
+
+        state["original_height"] = height
+        state["original_width"] = width
+        state["backbone_out"] = self.model.backbone.forward_image(image)
+        inst_interactivity_en = self.model.inst_interactive_predictor is not None
+        if inst_interactivity_en and "sam2_backbone_out" in state["backbone_out"]:
+            sam2_backbone_out = state["backbone_out"]["sam2_backbone_out"]
+            sam2_backbone_out["backbone_fpn"][0] = (
+                self.model.inst_interactive_predictor.model.sam_mask_decoder.conv_s0(
+                    sam2_backbone_out["backbone_fpn"][0]
+                )
+            )
+            sam2_backbone_out["backbone_fpn"][1] = (
+                self.model.inst_interactive_predictor.model.sam_mask_decoder.conv_s1(
+                    sam2_backbone_out["backbone_fpn"][1]
+                )
+            )
+        return state
+
+    @torch.inference_mode()
+    def set_image_batch(self, images: List[np.ndarray], state=None):
+        """Sets the image batch on which we want to do predictions."""
+        if state is None:
+            state = {}
+
+        if not isinstance(images, list):
+            raise ValueError("Images must be a list of PIL images or tensors")
+        assert len(images) > 0, "Images list must not be empty"
+        assert isinstance(
+            images[0], PIL.Image.Image
+        ), "Images must be a list of PIL images"
+
+        state["original_heights"] = [image.height for image in images]
+        state["original_widths"] = [image.width for image in images]
+
+        images = [
+            self.transform(v2.functional.to_image(image).to(self.device))
+            for image in images
+        ]
+        images = torch.stack(images, dim=0)
+        state["backbone_out"] = self.model.backbone.forward_image(images)
+        inst_interactivity_en = self.model.inst_interactive_predictor is not None
+        if inst_interactivity_en and "sam2_backbone_out" in state["backbone_out"]:
+            sam2_backbone_out = state["backbone_out"]["sam2_backbone_out"]
+            sam2_backbone_out["backbone_fpn"][0] = (
+                self.model.inst_interactive_predictor.model.sam_mask_decoder.conv_s0(
+                    sam2_backbone_out["backbone_fpn"][0]
+                )
+            )
+            sam2_backbone_out["backbone_fpn"][1] = (
+                self.model.inst_interactive_predictor.model.sam_mask_decoder.conv_s1(
+                    sam2_backbone_out["backbone_fpn"][1]
+                )
+            )
+        return state
+
+    @torch.inference_mode()
+    def set_text_prompt(self, prompt: str, state: Dict):
+        """Sets the text prompt and run the inference"""
+
+        if "backbone_out" not in state:
+            raise ValueError("You must call set_image before set_text_prompt")
+
+        text_outputs = self.model.backbone.forward_text([prompt], device=self.device)
+        # will erase the previous text prompt if any
+        state["backbone_out"].update(text_outputs)
+        if "geometric_prompt" not in state:
+            state["geometric_prompt"] = self.model._get_dummy_prompt()
+
+        return self._forward_grounding(state)
+
+    @torch.inference_mode()
+    def add_geometric_prompt(self, box: List, label: bool, state: Dict):
+        """Adds a box prompt and run the inference.
+        The image needs to be set, but not necessarily the text prompt.
+        The box is assumed to be in [center_x, center_y, width, height] format and normalized in [0, 1] range.
+        The label is True for a positive box, False for a negative box.
+        """
+        if "backbone_out" not in state:
+            raise ValueError("You must call set_image before set_text_prompt")
+
+        if "language_features" not in state["backbone_out"]:
+            # Looks like we don't have a text prompt yet. This is allowed, but we need to set the text prompt to "visual" for the model to rely only on the geometric prompt
+            dummy_text_outputs = self.model.backbone.forward_text(
+                ["visual"], device=self.device
+            )
+            state["backbone_out"].update(dummy_text_outputs)
+
+        if "geometric_prompt" not in state:
+            state["geometric_prompt"] = self.model._get_dummy_prompt()
+
+        # adding a batch and sequence dimension
+        boxes = torch.tensor(box, device=self.device, dtype=torch.float32).view(1, 1, 4)
+        labels = torch.tensor([label], device=self.device, dtype=torch.bool).view(1, 1)
+        state["geometric_prompt"].append_boxes(boxes, labels)
+
+        return self._forward_grounding(state)
+
+    def reset_all_prompts(self, state: Dict):
+        """Removes all the prompts and results"""
+        if "backbone_out" in state:
+            backbone_keys_to_del = [
+                "language_features",
+                "language_mask",
+                "language_embeds",
+            ]
+            for key in backbone_keys_to_del:
+                if key in state["backbone_out"]:
+                    del state["backbone_out"][key]
+
+        keys_to_del = ["geometric_prompt", "boxes", "masks", "masks_logits", "scores"]
+        for key in keys_to_del:
+            if key in state:
+                del state[key]
+
+    @torch.inference_mode()
+    def set_confidence_threshold(self, threshold: float, state=None):
+        """Sets the confidence threshold for the masks"""
+        self.confidence_threshold = threshold
+        if state is not None and "boxes" in state:
+            # we need to filter the boxes again
+            # In principle we could do this more efficiently since we would only need
+            # to rerun the heads. But this is simpler and not too inefficient
+            return self._forward_grounding(state)
+        return state
+
+    @torch.inference_mode()
+    def _forward_grounding(self, state: Dict):
+        outputs = self.model.forward_grounding(
+            backbone_out=state["backbone_out"],
+            find_input=self.find_stage,
+            geometric_prompt=state["geometric_prompt"],
+            find_target=None,
+        )
+
+        out_bbox = outputs["pred_boxes"]
+        out_logits = outputs["pred_logits"]
+        out_masks = outputs["pred_masks"]
+        out_probs = out_logits.sigmoid()
+        presence_score = outputs["presence_logit_dec"].sigmoid().unsqueeze(1)
+        out_probs = (out_probs * presence_score).squeeze(-1)
+
+        keep = out_probs > self.confidence_threshold
+        out_probs = out_probs[keep]
+        out_masks = out_masks[keep]
+        out_bbox = out_bbox[keep]
+
+        # convert to [x0, y0, x1, y1] format
+        boxes = box_ops.box_cxcywh_to_xyxy(out_bbox)
+
+        img_h = state["original_height"]
+        img_w = state["original_width"]
+        scale_fct = torch.tensor([img_w, img_h, img_w, img_h]).to(self.device)
+        boxes = boxes * scale_fct[None, :]
+
+        out_masks = interpolate(
+            out_masks.unsqueeze(1),
+            (img_h, img_w),
+            mode="bilinear",
+            align_corners=False,
+        ).sigmoid()
+
+        state["masks_logits"] = out_masks
+        state["masks"] = out_masks > 0.5
+        state["boxes"] = boxes
+        state["scores"] = out_probs
+        return state

sam3/model/sam3_tracker_utils.py

@@ -0,0 +1,427 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from numpy.typing import NDArray
+
+from sam3.model.edt import edt_triton
+
+
+def sample_box_points(
+    masks: torch.Tensor,
+    noise: float = 0.1,  # SAM default
+    noise_bound: int = 20,  # SAM default
+    top_left_label: int = 2,
+    bottom_right_label: int = 3,
+) -> tuple[NDArray, NDArray]:
+    """
+    Sample a noised version of the top left and bottom right corners of a given `bbox`
+
+    Inputs:
+    - masks: [B, 1, H, W] tensor
+    - noise: noise as a fraction of box width and height, dtype=float
+    - noise_bound: maximum amount of noise (in pure pixels), dtype=int
+
+    Returns:
+    - box_coords: [B, num_pt, 2], contains (x, y) coordinates of top left and bottom right box corners, dtype=torch.float
+    - box_labels: [B, num_pt], label 2 is reserverd for top left and 3 for bottom right corners, dtype=torch.int32
+    """
+    device = masks.device
+    box_coords = mask_to_box(masks)
+    B, _, H, W = masks.shape
+    box_labels = torch.tensor(
+        [top_left_label, bottom_right_label], dtype=torch.int, device=device
+    ).repeat(B)
+    if noise > 0.0:
+        if not isinstance(noise_bound, torch.Tensor):
+            noise_bound = torch.tensor(noise_bound, device=device)
+        bbox_w = box_coords[..., 2] - box_coords[..., 0]
+        bbox_h = box_coords[..., 3] - box_coords[..., 1]
+        max_dx = torch.min(bbox_w * noise, noise_bound)
+        max_dy = torch.min(bbox_h * noise, noise_bound)
+        box_noise = 2 * torch.rand(B, 1, 4, device=device) - 1
+        box_noise = box_noise * torch.stack((max_dx, max_dy, max_dx, max_dy), dim=-1)
+
+        box_coords = box_coords + box_noise
+        img_bounds = (
+            torch.tensor([W, H, W, H], device=device) - 1
+        )  # uncentered pixel coords
+        box_coords.clamp_(torch.zeros_like(img_bounds), img_bounds)  # In place clamping
+
+    box_coords = box_coords.reshape(-1, 2, 2)  # always 2 points
+    box_labels = box_labels.reshape(-1, 2)
+    return box_coords, box_labels
+
+
+def mask_to_box(masks: torch.Tensor):
+    """
+    compute bounding box given an input mask
+
+    Inputs:
+    - masks: [B, 1, H, W] tensor
+
+    Returns:
+    - box_coords: [B, 1, 4], contains (x, y) coordinates of top left and bottom right box corners, dtype=torch.Tensor
+    """
+    B, _, h, w = masks.shape
+    device = masks.device
+    mask_area = masks.sum(dim=(-1, -2))
+    xs = torch.arange(w, device=device, dtype=torch.int32)
+    ys = torch.arange(h, device=device, dtype=torch.int32)
+    grid_xs, grid_ys = torch.meshgrid(xs, ys, indexing="xy")
+    grid_xs = grid_xs[None, None, ...].expand(B, 1, h, w)
+    grid_ys = grid_ys[None, None, ...].expand(B, 1, h, w)
+    min_xs, _ = torch.min(torch.where(masks, grid_xs, w).flatten(-2), dim=-1)
+    max_xs, _ = torch.max(torch.where(masks, grid_xs, -1).flatten(-2), dim=-1)
+    min_ys, _ = torch.min(torch.where(masks, grid_ys, h).flatten(-2), dim=-1)
+    max_ys, _ = torch.max(torch.where(masks, grid_ys, -1).flatten(-2), dim=-1)
+    bbox_coords = torch.stack((min_xs, min_ys, max_xs, max_ys), dim=-1)
+    bbox_coords = torch.where(
+        mask_area[..., None] > 0, bbox_coords, torch.zeros_like(bbox_coords)
+    )
+    return bbox_coords
+
+
+def sample_random_points_from_errors(gt_masks, pred_masks, num_pt=1):
+    """
+    Sample `num_pt` random points (along with their labels) independently from the error regions.
+
+    Inputs:
+    - gt_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool
+    - pred_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool or None
+    - num_pt: int, number of points to sample independently for each of the B error maps
+
+    Outputs:
+    - points: [B, num_pt, 2], dtype=torch.float, contains (x, y) coordinates of each sampled point
+    - labels: [B, num_pt], dtype=torch.int32, where 1 means positive clicks and 0 means
+      negative clicks
+    """
+    if pred_masks is None:  # if pred_masks is not provided, treat it as empty
+        pred_masks = torch.zeros_like(gt_masks)
+    assert gt_masks.dtype == torch.bool and gt_masks.size(1) == 1
+    assert pred_masks.dtype == torch.bool and pred_masks.shape == gt_masks.shape
+    assert num_pt >= 0
+
+    B, _, H_im, W_im = gt_masks.shape
+    device = gt_masks.device
+
+    # false positive region, a new point sampled in this region should have
+    # negative label to correct the FP error
+    fp_masks = ~gt_masks & pred_masks
+    # false negative region, a new point sampled in this region should have
+    # positive label to correct the FN error
+    fn_masks = gt_masks & ~pred_masks
+    # whether the prediction completely match the ground-truth on each mask
+    all_correct = torch.all((gt_masks == pred_masks).flatten(2), dim=2)
+    all_correct = all_correct[..., None, None]
+
+    # channel 0 is FP map, while channel 1 is FN map
+    pts_noise = torch.rand(B, num_pt, H_im, W_im, 2, device=device)
+    # sample a negative new click from FP region or a positive new click
+    # from FN region, depend on where the maximum falls,
+    # and in case the predictions are all correct (no FP or FN), we just
+    # sample a negative click from the background region
+    pts_noise[..., 0] *= fp_masks | (all_correct & ~gt_masks)
+    pts_noise[..., 1] *= fn_masks
+    pts_idx = pts_noise.flatten(2).argmax(dim=2)
+    labels = (pts_idx % 2).to(torch.int32)
+    pts_idx = pts_idx // 2
+    pts_x = pts_idx % W_im
+    pts_y = pts_idx // W_im
+    points = torch.stack([pts_x, pts_y], dim=2).to(torch.float)
+    return points, labels
+
+
+def sample_one_point_from_error_center(gt_masks, pred_masks, padding=True):
+    """
+    Sample 1 random point (along with its label) from the center of each error region,
+    that is, the point with the largest distance to the boundary of each error region.
+    This is the RITM sampling method from https://github.com/saic-vul/ritm_interactive_segmentation/blob/master/isegm/inference/clicker.py
+
+    Inputs:
+    - gt_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool
+    - pred_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool or None
+    - padding: if True, pad with boundary of 1 px for distance transform
+
+    Outputs:
+    - points: [B, 1, 2], dtype=torch.float, contains (x, y) coordinates of each sampled point
+    - labels: [B, 1], dtype=torch.int32, where 1 means positive clicks and 0 means negative clicks
+    """
+    if pred_masks is None:
+        pred_masks = torch.zeros_like(gt_masks)
+    assert gt_masks.dtype == torch.bool and gt_masks.size(1) == 1
+    assert pred_masks.dtype == torch.bool and pred_masks.shape == gt_masks.shape
+
+    B, _, H, W = gt_masks.shape
+
+    # false positive region, a new point sampled in this region should have
+    # negative label to correct the FP error
+    fp_masks = (~gt_masks & pred_masks).squeeze(1)
+    # false negative region, a new point sampled in this region should have
+    # positive label to correct the FN error
+    fn_masks = (gt_masks & ~pred_masks).squeeze(1)
+
+    if padding:
+        padded_fp_masks = torch.zeros(
+            B, H + 2, W + 2, dtype=fp_masks.dtype, device=fp_masks.device
+        )
+        padded_fp_masks[:, 1 : H + 1, 1 : W + 1] = fp_masks
+        padded_fn_masks = torch.zeros(
+            B, H + 2, W + 2, dtype=fp_masks.dtype, device=fp_masks.device
+        )
+        padded_fn_masks[:, 1 : H + 1, 1 : W + 1] = fn_masks
+    else:
+        padded_fp_masks = fp_masks
+        padded_fn_masks = fn_masks
+
+    fn_mask_dt = edt_triton(padded_fn_masks)
+    fp_mask_dt = edt_triton(padded_fp_masks)
+    if padding:
+        fn_mask_dt = fn_mask_dt[:, 1:-1, 1:-1]
+        fp_mask_dt = fp_mask_dt[:, 1:-1, 1:-1]
+
+    fn_max, fn_argmax = fn_mask_dt.reshape(B, -1).max(dim=-1)
+    fp_max, fp_argmax = fp_mask_dt.reshape(B, -1).max(dim=-1)
+    is_positive = fn_max > fp_max
+    chosen = torch.where(is_positive, fn_argmax, fp_argmax)
+    points_x = chosen % W
+    points_y = chosen // W
+
+    labels = is_positive.long()
+    points = torch.stack([points_x, points_y], -1)
+    return points.unsqueeze(1), labels.unsqueeze(1)
+
+
+def sample_one_point_from_error_center_slow(gt_masks, pred_masks, padding=True):
+    """
+    Sample 1 random point (along with its label) from the center of each error region,
+    that is, the point with the largest distance to the boundary of each error region.
+    This is the RITM sampling method from https://github.com/saic-vul/ritm_interactive_segmentation/blob/master/isegm/inference/clicker.py
+
+    Inputs:
+    - gt_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool
+    - pred_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool or None
+    - padding: if True, pad with boundary of 1 px for distance transform
+
+    Outputs:
+    - points: [B, 1, 2], dtype=torch.float, contains (x, y) coordinates of each sampled point
+    - labels: [B, 1], dtype=torch.int32, where 1 means positive clicks and 0 means negative clicks
+    """
+    import cv2  # delay OpenCV import to avoid unnecessary dependency
+
+    if pred_masks is None:
+        pred_masks = torch.zeros_like(gt_masks)
+    assert gt_masks.dtype == torch.bool and gt_masks.size(1) == 1
+    assert pred_masks.dtype == torch.bool and pred_masks.shape == gt_masks.shape
+
+    B, _, _, W_im = gt_masks.shape
+    device = gt_masks.device
+
+    # false positive region, a new point sampled in this region should have
+    # negative label to correct the FP error
+    fp_masks = ~gt_masks & pred_masks
+    # false negative region, a new point sampled in this region should have
+    # positive label to correct the FN error
+    fn_masks = gt_masks & ~pred_masks
+
+    fp_masks = fp_masks.cpu().numpy()
+    fn_masks = fn_masks.cpu().numpy()
+    points = torch.zeros(B, 1, 2, dtype=torch.float)
+    labels = torch.ones(B, 1, dtype=torch.int32)
+    for b in range(B):
+        fn_mask = fn_masks[b, 0]
+        fp_mask = fp_masks[b, 0]
+        if padding:
+            fn_mask = np.pad(fn_mask, ((1, 1), (1, 1)), "constant")
+            fp_mask = np.pad(fp_mask, ((1, 1), (1, 1)), "constant")
+        # compute the distance of each point in FN/FP region to its boundary
+        fn_mask_dt = cv2.distanceTransform(fn_mask.astype(np.uint8), cv2.DIST_L2, 0)
+        fp_mask_dt = cv2.distanceTransform(fp_mask.astype(np.uint8), cv2.DIST_L2, 0)
+        if padding:
+            fn_mask_dt = fn_mask_dt[1:-1, 1:-1]
+            fp_mask_dt = fp_mask_dt[1:-1, 1:-1]
+
+        # take the point in FN/FP region with the largest distance to its boundary
+        fn_mask_dt_flat = fn_mask_dt.reshape(-1)
+        fp_mask_dt_flat = fp_mask_dt.reshape(-1)
+        fn_argmax = np.argmax(fn_mask_dt_flat)
+        fp_argmax = np.argmax(fp_mask_dt_flat)
+        is_positive = fn_mask_dt_flat[fn_argmax] > fp_mask_dt_flat[fp_argmax]
+        pt_idx = fn_argmax if is_positive else fp_argmax
+        points[b, 0, 0] = pt_idx % W_im  # x
+        points[b, 0, 1] = pt_idx // W_im  # y
+        labels[b, 0] = int(is_positive)
+
+    points = points.to(device)
+    labels = labels.to(device)
+    return points, labels
+
+
+def get_next_point(gt_masks, pred_masks, method):
+    if method == "uniform":
+        return sample_random_points_from_errors(gt_masks, pred_masks)
+    elif method == "center":
+        return sample_one_point_from_error_center(gt_masks, pred_masks)
+    else:
+        raise ValueError(f"unknown sampling method {method}")
+
+
+def select_closest_cond_frames(
+    frame_idx, cond_frame_outputs, max_cond_frame_num, keep_first_cond_frame=False
+):
+    """
+    Select up to `max_cond_frame_num` conditioning frames from `cond_frame_outputs`
+    that are temporally closest to the current frame at `frame_idx`. Here, we take
+    - a) the closest conditioning frame before `frame_idx` (if any);
+    - b) the closest conditioning frame after `frame_idx` (if any);
+    - c) any other temporally closest conditioning frames until reaching a total
+         of `max_cond_frame_num` conditioning frames.
+
+    Outputs:
+    - selected_outputs: selected items (keys & values) from `cond_frame_outputs`.
+    - unselected_outputs: items (keys & values) not selected in `cond_frame_outputs`.
+    """
+    if max_cond_frame_num == -1 or len(cond_frame_outputs) <= max_cond_frame_num:
+        selected_outputs = cond_frame_outputs
+        unselected_outputs = {}
+    else:
+        assert max_cond_frame_num >= 2, "we should allow using 2+ conditioning frames"
+        selected_outputs = {}
+        if keep_first_cond_frame:
+            idx_first = min(
+                (t for t in cond_frame_outputs if t < frame_idx), default=None
+            )
+            if idx_first is None:
+                # Maybe we are tracking in reverse
+                idx_first = max(
+                    (t for t in cond_frame_outputs if t > frame_idx), default=None
+                )
+            if idx_first is not None:
+                selected_outputs[idx_first] = cond_frame_outputs[idx_first]
+        # the closest conditioning frame before `frame_idx` (if any)
+        idx_before = max((t for t in cond_frame_outputs if t < frame_idx), default=None)
+        if idx_before is not None:
+            selected_outputs[idx_before] = cond_frame_outputs[idx_before]
+
+        # the closest conditioning frame after `frame_idx` (if any)
+        idx_after = min((t for t in cond_frame_outputs if t >= frame_idx), default=None)
+        if idx_after is not None:
+            selected_outputs[idx_after] = cond_frame_outputs[idx_after]
+
+        # add other temporally closest conditioning frames until reaching a total
+        # of `max_cond_frame_num` conditioning frames.
+        num_remain = max_cond_frame_num - len(selected_outputs)
+        inds_remain = sorted(
+            (t for t in cond_frame_outputs if t not in selected_outputs),
+            key=lambda x: abs(x - frame_idx),
+        )[:num_remain]
+        selected_outputs.update((t, cond_frame_outputs[t]) for t in inds_remain)
+        unselected_outputs = {
+            t: v for t, v in cond_frame_outputs.items() if t not in selected_outputs
+        }
+
+    return selected_outputs, unselected_outputs
+
+
+def get_1d_sine_pe(pos_inds, dim, temperature=10000):
+    """
+    Get 1D sine positional embedding as in the original Transformer paper.
+    """
+    pe_dim = dim // 2
+    dim_t = torch.arange(pe_dim, dtype=torch.float32, device=pos_inds.device)
+    dim_t = temperature ** (2 * (dim_t // 2) / pe_dim)
+
+    pos_embed = pos_inds.unsqueeze(-1) / dim_t
+    pos_embed = torch.cat([pos_embed.sin(), pos_embed.cos()], dim=-1)
+    return pos_embed
+
+
+def get_best_gt_match_from_multimasks(pred_multimasks, gt_masks, pred_scores=None):
+    """
+    Get the mask with the best match to GT masks (based on IoU) from pred_multimasks.
+    Optionally, use `pred_scores` to break ties in case all IoUs are zeros.
+    """
+    assert pred_multimasks.ndim == 4 and gt_masks.ndim == 4
+    if pred_multimasks.size(1) == 1:
+        return pred_multimasks  # only a single mask channel, nothing to select
+
+    pred_multimasks_binary = pred_multimasks > 0
+    area_i = torch.sum(pred_multimasks_binary & gt_masks, dim=(2, 3)).float()
+    area_u = torch.sum(pred_multimasks_binary | gt_masks, dim=(2, 3)).float()
+    ious = area_i / torch.clamp(area_u, min=1.0)
+
+    # In case all IoUs are zeros (e.g. because the GT mask is empty), use pred_scores
+    # to break ties and select the best mask
+    if pred_scores is not None:
+        has_nonzero_ious = torch.any(ious > 0).expand_as(ious)
+        scores = torch.where(has_nonzero_ious, ious, pred_scores)
+    else:
+        scores = ious
+
+    # Finally, take the best mask prediction (with the highest score)
+    best_scores_inds = torch.argmax(scores, dim=-1)
+    batch_inds = torch.arange(scores.size(0), device=scores.device)
+    best_pred_mask = pred_multimasks[batch_inds, best_scores_inds].unsqueeze(1)
+    return best_pred_mask
+
+
+def fill_holes_in_mask_scores(mask, max_area, fill_holes=True, remove_sprinkles=True):
+    """
+    A post processor to fill small holes in mask scores with area under `max_area`.
+    Holes are those small connected components in either background or foreground.
+
+    Note that it relies on the "cc_torch" package to find connected components fast. You can
+    install it via the following command (`TORCH_CUDA_ARCH_LIST=8.0` is for A100 GPUs):
+    ```
+    pip uninstall -y cc_torch; TORCH_CUDA_ARCH_LIST=8.0 9.0 pip install git+https://github.com/ronghanghu/cc_torch
+    ```
+    Otherwise, it will fallback to a slightly slower triton implementation, or skimage if the tensor is on cpu
+    """
+
+    if max_area <= 0:
+        return mask  # nothing to fill in this case
+
+    if fill_holes:
+        # We remove small connected components in background by changing them to foreground
+        # with a small positive mask score (0.1).
+        mask_bg = mask <= 0
+        bg_area_thresh = max_area
+        _, areas_bg = _get_connected_components_with_padding(mask_bg)
+        small_components_bg = mask_bg & (areas_bg <= bg_area_thresh)
+        mask = torch.where(small_components_bg, 0.1, mask)
+
+    if remove_sprinkles:
+        # We remove small connected components in foreground by changing them to background
+        # with a small negative mask score (-0.1). Here we only remove connected components
+        # whose areas are under both `max_area` and half of the entire mask's area. This
+        # removes sprinkles while avoids filtering out tiny objects that we want to track.
+        mask_fg = mask > 0
+        fg_area_thresh = torch.sum(mask_fg, dim=(2, 3), keepdim=True, dtype=torch.int32)
+        fg_area_thresh.floor_divide_(2).clamp_(max=max_area)
+        _, areas_fg = _get_connected_components_with_padding(mask_fg)
+        small_components_fg = mask_fg & (areas_fg <= fg_area_thresh)
+        mask = torch.where(small_components_fg, -0.1, mask)
+    return mask
+
+
+def _get_connected_components_with_padding(mask):
+    """Get connected components from masks (possibly padding them to an even size)."""
+    from sam3.perflib.connected_components import connected_components
+
+    mask = mask.to(torch.uint8)
+    _, _, H, W = mask.shape
+    # make sure both height and width are even (to be compatible with cc_torch)
+    pad_h = H % 2
+    pad_w = W % 2
+    if pad_h == 0 and pad_w == 0:
+        labels, counts = connected_components(mask)
+    else:
+        # pad the mask to make its height and width even
+        # padding format is (padding_left,padding_right,padding_top,padding_bottom)
+        mask_pad = F.pad(mask, (0, pad_w, 0, pad_h), mode="constant", value=0)
+        labels, counts = connected_components(mask_pad)
+        labels = labels[:, :, :H, :W]
+        counts = counts[:, :, :H, :W]
+
+    return labels, counts

sam3/model/sam3_video_predictor.py

@@ -0,0 +1,521 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import datetime
+import gc
+import multiprocessing as mp
+import os
+import queue
+import socket
+import sys
+import time
+import uuid
+from contextlib import closing
+from typing import List, Optional
+
+import psutil
+import torch
+
+from sam3.logger import get_logger
+
+logger = get_logger(__name__)
+
+
+class Sam3VideoPredictor:
+    # a global dictionary that holds all inference states for this model (key is session_id)
+    _ALL_INFERENCE_STATES = {}
+
+    def __init__(
+        self,
+        checkpoint_path=None,
+        bpe_path=None,
+        has_presence_token=True,
+        geo_encoder_use_img_cross_attn=True,
+        strict_state_dict_loading=True,
+        async_loading_frames=False,
+        video_loader_type="cv2",
+        apply_temporal_disambiguation: bool = True,
+    ):
+        self.async_loading_frames = async_loading_frames
+        self.video_loader_type = video_loader_type
+        from sam3.model_builder import build_sam3_video_model
+
+        self.model = (
+            build_sam3_video_model(
+                checkpoint_path=checkpoint_path,
+                bpe_path=bpe_path,
+                has_presence_token=has_presence_token,
+                geo_encoder_use_img_cross_attn=geo_encoder_use_img_cross_attn,
+                strict_state_dict_loading=strict_state_dict_loading,
+                apply_temporal_disambiguation=apply_temporal_disambiguation,
+            )
+            .cuda()
+            .eval()
+        )
+
+    @torch.inference_mode()
+    def handle_request(self, request):
+        """Dispatch a request based on its type."""
+        request_type = request["type"]
+        if request_type == "start_session":
+            return self.start_session(
+                resource_path=request["resource_path"],
+                session_id=request.get("session_id", None),
+            )
+        elif request_type == "add_prompt":
+            return self.add_prompt(
+                session_id=request["session_id"],
+                frame_idx=request["frame_index"],
+                text=request.get("text", None),
+                points=request.get("points", None),
+                point_labels=request.get("point_labels", None),
+                bounding_boxes=request.get("bounding_boxes", None),
+                bounding_box_labels=request.get("bounding_box_labels", None),
+                obj_id=request.get("obj_id", None),
+            )
+        elif request_type == "remove_object":
+            return self.remove_object(
+                session_id=request["session_id"],
+                obj_id=request["obj_id"],
+                is_user_action=request.get("is_user_action", True),
+            )
+        elif request_type == "reset_session":
+            return self.reset_session(session_id=request["session_id"])
+        elif request_type == "close_session":
+            return self.close_session(session_id=request["session_id"])
+        else:
+            raise RuntimeError(f"invalid request type: {request_type}")
+
+    @torch.inference_mode()
+    def handle_stream_request(self, request):
+        """Dispatch a stream request based on its type."""
+        request_type = request["type"]
+        if request_type == "propagate_in_video":
+            yield from self.propagate_in_video(
+                session_id=request["session_id"],
+                propagation_direction=request.get("propagation_direction", "both"),
+                start_frame_idx=request.get("start_frame_index", None),
+                max_frame_num_to_track=request.get("max_frame_num_to_track", None),
+            )
+        else:
+            raise RuntimeError(f"invalid request type: {request_type}")
+
+    def start_session(self, resource_path, session_id=None):
+        """
+        Start a new inference session on an image or a video. Here `resource_path`
+        can be either a path to an image file (for image inference) or an MP4 file
+        or directory with JPEG video frames (for video inference).
+
+        If `session_id` is defined, it will be used as identifier for the
+        session. If it is not defined, the start_session function will create
+        a session id and return it.
+        """
+        # get an initial inference_state from the model
+        inference_state = self.model.init_state(
+            resource_path=resource_path,
+            async_loading_frames=self.async_loading_frames,
+            video_loader_type=self.video_loader_type,
+        )
+        if not session_id:
+            session_id = str(uuid.uuid4())
+        self._ALL_INFERENCE_STATES[session_id] = {
+            "state": inference_state,
+            "session_id": session_id,
+            "start_time": time.time(),
+        }
+        logger.debug(
+            f"started new session {session_id}; {self._get_session_stats()}; "
+            f"{self._get_torch_and_gpu_properties()}"
+        )
+        return {"session_id": session_id}
+
+    def add_prompt(
+        self,
+        session_id: str,
+        frame_idx: int,
+        text: Optional[str] = None,
+        points: Optional[List[List[float]]] = None,
+        point_labels: Optional[List[int]] = None,
+        bounding_boxes: Optional[List[List[float]]] = None,
+        bounding_box_labels: Optional[List[int]] = None,
+        obj_id: Optional[int] = None,
+    ):
+        """Add text, box and/or point prompt on a specific video frame."""
+        logger.debug(
+            f"add prompt on frame {frame_idx} in session {session_id}: "
+            f"{text=}, {points=}, {point_labels=}, "
+            f"{bounding_boxes=}, {bounding_box_labels=}"
+        )
+        session = self._get_session(session_id)
+        inference_state = session["state"]
+
+        frame_idx, outputs = self.model.add_prompt(
+            inference_state=inference_state,
+            frame_idx=frame_idx,
+            text_str=text,
+            points=points,
+            point_labels=point_labels,
+            boxes_xywh=bounding_boxes,
+            box_labels=bounding_box_labels,
+            obj_id=obj_id,
+        )
+        return {"frame_index": frame_idx, "outputs": outputs}
+
+    def remove_object(
+        self,
+        session_id: str,
+        obj_id: int,
+        is_user_action: bool = True,
+    ):
+        """Remove an object from tracking."""
+        logger.debug(
+            f"remove object {obj_id} in session {session_id}: " f"{is_user_action=}"
+        )
+        session = self._get_session(session_id)
+        inference_state = session["state"]
+
+        self.model.remove_object(
+            inference_state=inference_state,
+            obj_id=obj_id,
+            is_user_action=is_user_action,
+        )
+        return {"is_success": True}
+
+    def propagate_in_video(
+        self,
+        session_id,
+        propagation_direction,
+        start_frame_idx,
+        max_frame_num_to_track,
+    ):
+        """Propagate the added prompts to get grounding results on all video frames."""
+        logger.debug(
+            f"propagate in video in session {session_id}: "
+            f"{propagation_direction=}, {start_frame_idx=}, {max_frame_num_to_track=}"
+        )
+        try:
+            session = self._get_session(session_id)
+            inference_state = session["state"]
+            if propagation_direction not in ["both", "forward", "backward"]:
+                raise ValueError(
+                    f"invalid propagation direction: {propagation_direction}"
+                )
+
+            # First doing the forward propagation
+            if propagation_direction in ["both", "forward"]:
+                for frame_idx, outputs in self.model.propagate_in_video(
+                    inference_state=inference_state,
+                    start_frame_idx=start_frame_idx,
+                    max_frame_num_to_track=max_frame_num_to_track,
+                    reverse=False,
+                ):
+                    yield {"frame_index": frame_idx, "outputs": outputs}
+            # Then doing the backward propagation (reverse in time)
+            if propagation_direction in ["both", "backward"]:
+                for frame_idx, outputs in self.model.propagate_in_video(
+                    inference_state=inference_state,
+                    start_frame_idx=start_frame_idx,
+                    max_frame_num_to_track=max_frame_num_to_track,
+                    reverse=True,
+                ):
+                    yield {"frame_index": frame_idx, "outputs": outputs}
+        finally:
+            # Log upon completion (so that e.g. we can see if two propagations happen in parallel).
+            # Using `finally` here to log even when the tracking is aborted with GeneratorExit.
+            logger.debug(
+                f"propagation ended in session {session_id}; {self._get_session_stats()}"
+            )
+
+    def reset_session(self, session_id):
+        """Reset the session to its initial state (as when it's initial opened)."""
+        logger.debug(f"reset session {session_id}")
+        session = self._get_session(session_id)
+        inference_state = session["state"]
+        self.model.reset_state(inference_state)
+        return {"is_success": True}
+
+    def close_session(self, session_id):
+        """
+        Close a session. This method is idempotent and can be called multiple
+        times on the same "session_id".
+        """
+        session = self._ALL_INFERENCE_STATES.pop(session_id, None)
+        if session is None:
+            logger.warning(
+                f"cannot close session {session_id} as it does not exist (it might have expired); "
+                f"{self._get_session_stats()}"
+            )
+        else:
+            del session
+            gc.collect()
+            logger.info(f"removed session {session_id}; {self._get_session_stats()}")
+        return {"is_success": True}
+
+    def _get_session(self, session_id):
+        session = self._ALL_INFERENCE_STATES.get(session_id, None)
+        if session is None:
+            raise RuntimeError(
+                f"Cannot find session {session_id}; it might have expired"
+            )
+        return session
+
+    def _get_session_stats(self):
+        """Get a statistics string for live sessions and their GPU usage."""
+        # print both the session ids and their video frame numbers
+        live_session_strs = [
+            f"'{session_id}' ({session['state']['num_frames']} frames)"
+            for session_id, session in self._ALL_INFERENCE_STATES.items()
+        ]
+        session_stats_str = (
+            f"live sessions: [{', '.join(live_session_strs)}], GPU memory: "
+            f"{torch.cuda.memory_allocated() // 1024**2} MiB used and "
+            f"{torch.cuda.memory_reserved() // 1024**2} MiB reserved"
+            f" (max over time: {torch.cuda.max_memory_allocated() // 1024**2} MiB used "
+            f"and {torch.cuda.max_memory_reserved() // 1024**2} MiB reserved)"
+        )
+        return session_stats_str
+
+    def _get_torch_and_gpu_properties(self):
+        """Get a string for PyTorch and GPU properties (for logging and debugging)."""
+        torch_and_gpu_str = (
+            f"torch: {torch.__version__} with CUDA arch {torch.cuda.get_arch_list()}, "
+            f"GPU device: {torch.cuda.get_device_properties(torch.cuda.current_device())}"
+        )
+        return torch_and_gpu_str
+
+    def shutdown(self):
+        """Shutdown the predictor and clear all sessions."""
+        self._ALL_INFERENCE_STATES.clear()
+
+
+class Sam3VideoPredictorMultiGPU(Sam3VideoPredictor):
+    def __init__(self, *model_args, gpus_to_use=None, **model_kwargs):
+        if gpus_to_use is None:
+            # if not specified, use only the current GPU by default
+            gpus_to_use = [torch.cuda.current_device()]
+
+        IS_MAIN_PROCESS = os.getenv("IS_MAIN_PROCESS", "1") == "1"
+        if IS_MAIN_PROCESS:
+            gpus_to_use = sorted(set(gpus_to_use))
+            logger.info(f"using the following GPU IDs: {gpus_to_use}")
+            assert len(gpus_to_use) > 0 and all(isinstance(i, int) for i in gpus_to_use)
+            assert all(0 <= i < torch.cuda.device_count() for i in gpus_to_use)
+            os.environ["MASTER_ADDR"] = "localhost"
+            os.environ["MASTER_PORT"] = f"{self._find_free_port()}"
+            os.environ["RANK"] = "0"
+            os.environ["WORLD_SIZE"] = f"{len(gpus_to_use)}"
+
+        self.gpus_to_use = gpus_to_use
+        self.rank = int(os.environ["RANK"])
+        self.world_size = int(os.environ["WORLD_SIZE"])
+        self.rank_str = f"rank={self.rank} with world_size={self.world_size}"
+        self.device = torch.device(f"cuda:{self.gpus_to_use[self.rank]}")
+        torch.cuda.set_device(self.device)
+        self.has_shutdown = False
+        if self.rank == 0:
+            logger.info("\n\n\n\t*** START loading model on all ranks ***\n\n")
+
+        logger.info(f"loading model on {self.rank_str} -- this could take a while ...")
+        super().__init__(*model_args, **model_kwargs)
+        logger.info(f"loading model on {self.rank_str} -- DONE locally")
+
+        if self.world_size > 1 and self.rank == 0:
+            # start the worker processes *after* the model is loaded in the main process
+            # so that the main process can run torch.compile and fill the cache first
+            self._start_worker_processes(*model_args, **model_kwargs)
+            for rank in range(1, self.world_size):
+                self.command_queues[rank].put(("start_nccl_process_group", None))
+            self._start_nccl_process_group()
+
+        if self.rank == 0:
+            logger.info("\n\n\n\t*** DONE loading model on all ranks ***\n\n")
+
+    @torch.inference_mode()
+    def handle_request(self, request):
+        """Dispatch a request based on its type."""
+        if self.has_shutdown:
+            raise RuntimeError(
+                "cannot handle request after the predictor has shutdown; please create a new predictor"
+            )
+
+        # when starting a session, we need to create a session id before dispatching
+        # the request to the workers
+        if request["type"] == "start_session" and request.get("session_id") is None:
+            request["session_id"] = str(uuid.uuid4())
+        # dispatch the request to all worker processes
+        if self.world_size > 1 and self.rank == 0:
+            for rank in range(1, self.world_size):
+                self.command_queues[rank].put((request, False))
+
+        response = super().handle_request(request)
+
+        if self.world_size > 1:
+            torch.distributed.barrier()  # wait for all ranks to finish
+        return response
+
+    @torch.inference_mode()
+    def handle_stream_request(self, request):
+        """Dispatch a stream request based on its type."""
+        if self.has_shutdown:
+            raise RuntimeError(
+                "cannot handle request after the predictor has shutdown; please create a new predictor"
+            )
+
+        # dispatch the request to all worker processes
+        if self.world_size > 1 and self.rank == 0:
+            for rank in range(1, self.world_size):
+                self.command_queues[rank].put((request, True))
+
+        yield from super().handle_stream_request(request)
+
+        if self.world_size > 1:
+            torch.distributed.barrier()  # wait for all ranks to finish
+
+    def _start_worker_processes(self, *model_args, **model_kwargs):
+        """Start worker processes for handling model inference."""
+        world_size = self.world_size
+        logger.info(f"spawning {world_size - 1} worker processes")
+        # Use "spawn" (instead of "fork") for different PyTorch or CUDA context
+        mp_ctx = mp.get_context("spawn")
+        self.command_queues = {rank: mp_ctx.Queue() for rank in range(1, world_size)}
+        self.result_queues = {rank: mp_ctx.Queue() for rank in range(1, world_size)}
+        parent_pid = os.getpid()
+        for rank in range(1, world_size):
+            # set the environment variables for each worker process
+            os.environ["IS_MAIN_PROCESS"] = "0"  # mark this as a worker process
+            os.environ["RANK"] = f"{rank}"
+            worker_process = mp_ctx.Process(
+                target=Sam3VideoPredictorMultiGPU._worker_process_command_loop,
+                args=(
+                    rank,
+                    world_size,
+                    self.command_queues[rank],
+                    self.result_queues[rank],
+                    model_args,
+                    model_kwargs,
+                    self.gpus_to_use,
+                    parent_pid,
+                ),
+                daemon=True,
+            )
+            worker_process.start()
+        # revert the environment variables for the main process
+        os.environ["IS_MAIN_PROCESS"] = "1"
+        os.environ["RANK"] = "0"
+        # wait for all the worker processes to load the model and collect their PIDs
+        self.worker_pids = {}
+        for rank in range(1, self.world_size):
+            # a large timeout to cover potentially long model loading time due to compilation
+            _, worker_pid = self.result_queues[rank].get(timeout=7200)
+            self.worker_pids[rank] = worker_pid
+        logger.info(f"spawned {world_size - 1} worker processes")
+
+    def _start_nccl_process_group(self):
+        rank = int(os.environ["RANK"])
+        world_size = int(os.environ["WORLD_SIZE"])
+        if world_size == 1:
+            return
+
+        logger.debug(f"starting NCCL process group on {rank=} with {world_size=}")
+        assert not torch.distributed.is_initialized()
+        # use the "env://" init method with environment variables set in start_worker_processes
+        # a short 3-min timeout to quickly detect any synchronization failures
+        timeout_sec = int(os.getenv("SAM3_COLLECTIVE_OP_TIMEOUT_SEC", "180"))
+        timeout = datetime.timedelta(seconds=timeout_sec)
+        torch.distributed.init_process_group(
+            backend="nccl",
+            init_method="env://",
+            timeout=timeout,
+            device_id=self.device,
+        )
+        # warm-up the NCCL process group by running a dummy all-reduce
+        tensor = torch.ones(1024, 1024).cuda()
+        torch.distributed.all_reduce(tensor)
+        logger.debug(f"started NCCL process group on {rank=} with {world_size=}")
+
+    def _find_free_port(self) -> int:
+        """
+        Find a free port (a random free port from 1024 to 65535 will be selected)
+        https://stackoverflow.com/questions/1365265/on-localhost-how-do-i-pick-a-free-port-number)
+        """
+        with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as s:
+            s.bind(("", 0))
+            s.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
+            return s.getsockname()[1]
+
+    @staticmethod
+    def _worker_process_command_loop(
+        rank,
+        world_size,
+        command_queue,
+        result_queue,
+        model_args,
+        model_kwargs,
+        gpus_to_use,
+        parent_pid,
+    ):
+        """
+        The command loop for each worker process. It listens to commands from the main process
+        and executes them using the model.
+        """
+        logger.info(f"starting worker process {rank=} with {world_size=}")
+        # verify that the environment variables are set correctly
+        assert int(os.environ["IS_MAIN_PROCESS"]) == 0
+        assert int(os.environ["RANK"]) == rank
+        assert int(os.environ["WORLD_SIZE"]) == world_size
+        # load the model in this worker process
+        predictor = Sam3VideoPredictorMultiGPU(
+            *model_args, gpus_to_use=gpus_to_use, **model_kwargs
+        )
+        logger.info(f"started worker {rank=} with {world_size=}")
+        # return the worker process id to the main process for bookkeeping
+        worker_pid = os.getpid()
+        result_queue.put(("load_model", worker_pid))
+
+        # wait for the command to start the NCCL process group
+        request_type, _ = command_queue.get(timeout=7200)
+        assert request_type == "start_nccl_process_group"
+        predictor._start_nccl_process_group()
+
+        # keep listening to commands from the main process
+        while True:
+            try:
+                request, is_stream_request = command_queue.get(timeout=5.0)
+                if request == "shutdown":
+                    logger.info(f"worker {rank=} shutting down")
+                    torch.distributed.destroy_process_group()
+                    result_queue.put(("shutdown", True))  # acknowledge the shutdown
+                    sys.exit(0)
+
+                logger.debug(f"worker {rank=} received request {request['type']=}")
+                if is_stream_request:
+                    for _ in predictor.handle_stream_request(request):
+                        pass  # handle stream requests in a generator fashion
+                else:
+                    predictor.handle_request(request)
+            except queue.Empty:
+                # Usually Python's multiprocessing module will shutdown all the daemon worker
+                # processes when the main process exits gracefully. However, the user may kill
+                # the main process using SIGKILL and thereby leaving no chance for the main process
+                # to clean up its daemon child processes. So here we manually check whether the
+                # parent process still exists (every 5 sec as in `command_queue.get` timeout).
+                if not psutil.pid_exists(parent_pid):
+                    logger.info(
+                        f"stopping worker {rank=} as its parent process has exited"
+                    )
+                    sys.exit(1)
+            except Exception as e:
+                logger.error(f"worker {rank=} exception: {e}", exc_info=True)
+
+    def shutdown(self):
+        """Shutdown all worker processes."""
+        if self.rank == 0 and self.world_size > 1:
+            logger.info(f"shutting down {self.world_size - 1} worker processes")
+            for rank in range(1, self.world_size):
+                self.command_queues[rank].put(("shutdown", False))
+            torch.distributed.destroy_process_group()
+            for rank in range(1, self.world_size):
+                self.result_queues[rank].get()  # wait for the worker to acknowledge
+            logger.info(f"shut down {self.world_size - 1} worker processes")
+        self.has_shutdown = True
+
+        super().shutdown()

sam3/model/text_encoder_ve.py

@@ -0,0 +1,328 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+from collections import OrderedDict
+from typing import Callable, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+from torch.utils.checkpoint import checkpoint
+
+from .model_misc import LayerScale
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        n_head: int,
+        mlp_ratio: float = 4.0,
+        ls_init_value: Optional[float] = None,
+        act_layer: Callable[[], nn.Module] = nn.GELU,
+        norm_layer: Callable[[int], nn.Module] = nn.LayerNorm,
+    ):
+        super().__init__()
+        # Attention
+        self.attn = nn.MultiheadAttention(d_model, n_head, batch_first=True)
+
+        # LayerNorm, LayerScale
+        self.ln_1 = norm_layer(d_model)
+        self.ln_2 = norm_layer(d_model)
+
+        self.ls_1 = (
+            LayerScale(d_model, ls_init_value)
+            if ls_init_value is not None
+            else nn.Identity()
+        )
+        self.ls_2 = (
+            LayerScale(d_model, ls_init_value)
+            if ls_init_value is not None
+            else nn.Identity()
+        )
+
+        # MLP
+        mlp_width = int(d_model * mlp_ratio)
+        self.mlp = nn.Sequential(
+            OrderedDict(
+                [
+                    ("c_fc", nn.Linear(d_model, mlp_width)),
+                    ("gelu", act_layer()),
+                    ("c_proj", nn.Linear(mlp_width, d_model)),
+                ]
+            )
+        )
+
+    def attention(
+        self,
+        q_x: torch.Tensor,
+        k_x: Optional[torch.Tensor] = None,
+        v_x: Optional[torch.Tensor] = None,
+        attn_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        k_x = k_x if k_x is not None else q_x
+        v_x = v_x if v_x is not None else q_x
+        if attn_mask is not None:
+            # Leave boolean masks as is
+            if not attn_mask.dtype == torch.bool:
+                attn_mask = attn_mask.to(q_x.dtype)
+
+        return self.attn(q_x, k_x, v_x, need_weights=False, attn_mask=attn_mask)[0]
+
+    def forward(
+        self,
+        q_x: torch.Tensor,
+        k_x: Optional[torch.Tensor] = None,
+        v_x: Optional[torch.Tensor] = None,
+        attn_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        k_x = (
+            self.ln_1_kv(k_x) if hasattr(self, "ln_1_kv") and k_x is not None else None
+        )
+        v_x = (
+            self.ln_1_kv(v_x) if hasattr(self, "ln_1_kv") and v_x is not None else None
+        )
+        x = q_x + self.ls_1(
+            self.attention(q_x=self.ln_1(q_x), k_x=k_x, v_x=v_x, attn_mask=attn_mask)
+        )
+        x = x + self.ls_2(self.mlp(self.ln_2(x)))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        width: int,
+        layers: int,
+        heads: int,
+        mlp_ratio: float = 4.0,
+        ls_init_value: Optional[float] = None,
+        act_layer: Callable[[], nn.Module] = nn.GELU,
+        norm_layer: Callable[[int], nn.Module] = nn.LayerNorm,
+        compile_mode: Optional[str] = None,
+        use_act_checkpoint: bool = False,
+    ):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.grad_checkpointing = use_act_checkpoint
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(
+                    width,
+                    heads,
+                    mlp_ratio,
+                    ls_init_value=ls_init_value,
+                    act_layer=act_layer,
+                    norm_layer=norm_layer,
+                )
+                for _ in range(layers)
+            ]
+        )
+
+        if compile_mode is not None:
+            self.forward = torch.compile(
+                self.forward, mode=compile_mode, fullgraph=True
+            )
+            if self.grad_checkpointing:
+                torch._dynamo.config.optimize_ddp = False
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        attn_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        for _, r in enumerate(self.resblocks):
+            if (
+                self.grad_checkpointing
+                and not torch.jit.is_scripting()
+                and self.training
+            ):
+                x = checkpoint(r, x, None, None, attn_mask, use_reentrant=False)
+            else:
+                x = r(
+                    x,
+                    attn_mask=attn_mask,
+                )
+        return x
+
+
+def text_global_pool(
+    x: torch.Tensor, text: Optional[torch.Tensor] = None, pool_type: str = "argmax"
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    if pool_type == "first":
+        pooled, tokens = x[:, 0], x[:, 1:]
+    elif pool_type == "last":
+        pooled, tokens = x[:, -1], x[:, :-1]
+    elif pool_type == "argmax":
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        assert text is not None
+        pooled, tokens = x[torch.arange(x.shape[0]), text.argmax(dim=-1)], x
+    else:
+        pooled = tokens = x
+    return pooled, tokens
+
+
+class TextTransformer(nn.Module):
+    def __init__(
+        self,
+        context_length: int = 77,
+        vocab_size: int = 49408,
+        width: int = 512,
+        heads: int = 8,
+        layers: int = 12,
+        mlp_ratio: float = 4.0,
+        ls_init_value: Optional[float] = None,
+        output_dim: int = 512,
+        no_causal_mask: bool = False,
+        pool_type: str = "none",  # no pooling
+        proj_bias: bool = False,
+        act_layer: Callable = nn.GELU,
+        norm_layer: Callable = nn.LayerNorm,
+        output_tokens: bool = False,
+        use_ln_post: bool = True,
+        compile_mode: Optional[str] = None,
+        use_act_checkpoint: bool = False,
+    ):
+        super().__init__()
+        assert pool_type in ("first", "last", "argmax", "none")
+        self.output_tokens = output_tokens
+        self.num_pos = self.context_length = context_length
+        self.vocab_size = vocab_size
+        self.width = width
+        self.output_dim = output_dim
+        self.heads = heads
+        self.pool_type = pool_type
+
+        self.token_embedding = nn.Embedding(self.vocab_size, width)
+        self.positional_embedding = nn.Parameter(torch.empty(self.num_pos, width))
+        self.transformer = Transformer(
+            width=width,
+            layers=layers,
+            heads=heads,
+            mlp_ratio=mlp_ratio,
+            ls_init_value=ls_init_value,
+            act_layer=act_layer,
+            norm_layer=norm_layer,
+            compile_mode=compile_mode,
+            use_act_checkpoint=use_act_checkpoint,
+        )
+        self.ln_final = norm_layer(width) if use_ln_post else nn.Identity()
+        if no_causal_mask:
+            self.attn_mask = None
+        else:
+            self.register_buffer(
+                "attn_mask", self.build_causal_mask(), persistent=False
+            )
+        if proj_bias:
+            self.text_projection = nn.Linear(width, output_dim)
+        else:
+            self.text_projection = nn.Parameter(torch.empty(width, output_dim))
+
+    def build_causal_mask(self) -> torch.Tensor:
+        # lazily create causal attention mask, with full attention between the tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(self.num_pos, self.num_pos)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)  # zero out the lower diagonal
+        return mask
+
+    def forward(
+        self, text: torch.Tensor
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        seq_len = text.shape[1]
+        x = self.token_embedding(text)  # [batch_size, n_ctx, d_model]
+
+        attn_mask = self.attn_mask
+        if attn_mask is not None:
+            attn_mask = attn_mask[:seq_len, :seq_len]
+
+        x = x + self.positional_embedding[:seq_len]
+        x = self.transformer(x, attn_mask=attn_mask)
+
+        x = self.ln_final(x)
+        pooled, tokens = text_global_pool(x, text, pool_type=self.pool_type)
+        if self.text_projection is not None:
+            if isinstance(self.text_projection, nn.Linear):
+                pooled = self.text_projection(pooled)
+            else:
+                pooled = pooled @ self.text_projection
+        if self.output_tokens:
+            return pooled, tokens
+        return pooled
+
+
+class VETextEncoder(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        tokenizer: Callable,
+        width: int = 1024,
+        heads: int = 16,
+        layers: int = 24,
+        context_length: int = 32,
+        vocab_size: int = 49408,
+        use_ln_post: bool = True,
+        compile_mode: Optional[str] = None,
+        use_act_checkpoint: bool = True,
+    ):
+        super().__init__()
+        self.context_length = context_length
+        self.use_ln_post = use_ln_post
+        self.tokenizer = tokenizer
+
+        self.encoder = TextTransformer(
+            context_length=self.context_length,
+            vocab_size=vocab_size,
+            width=width,
+            heads=heads,
+            layers=layers,
+            # we want the tokens, not just the pooled output
+            output_tokens=True,
+            use_ln_post=use_ln_post,
+            compile_mode=compile_mode,
+            use_act_checkpoint=use_act_checkpoint,
+        )
+        self.resizer = nn.Linear(self.encoder.width, d_model)
+
+    def forward(
+        self,
+        text: Union[List[str], Tuple[torch.Tensor, torch.Tensor, dict]],
+        input_boxes: Optional[List] = None,
+        device: torch.device = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        if isinstance(text[0], str):
+            # no use case for this
+            assert input_boxes is None or len(input_boxes) == 0, "not supported"
+
+            # Encode the text
+            tokenized = self.tokenizer(text, context_length=self.context_length).to(
+                device
+            )  # [b, seq_len]
+            text_attention_mask = (tokenized != 0).bool()
+
+            # manually embed the tokens
+            inputs_embeds = self.encoder.token_embedding(
+                tokenized
+            )  # [b, seq_len, d=1024]
+            _, text_memory = self.encoder(tokenized)  # [b, seq_len, d=1024]
+
+            assert text_memory.shape[1] == inputs_embeds.shape[1]
+            # Invert attention mask because its the opposite in pytorch transformer
+            text_attention_mask = text_attention_mask.ne(1)
+            # Transpose memory because pytorch's attention expects sequence first
+            text_memory = text_memory.transpose(0, 1)
+            # Resize the encoder hidden states to be of the same d_model as the decoder
+            text_memory_resized = self.resizer(text_memory)
+        else:
+            # The text is already encoded, use as is.
+            text_attention_mask, text_memory_resized, tokenized = text
+            inputs_embeds = tokenized["inputs_embeds"]
+            assert (
+                input_boxes is None or len(input_boxes) == 0
+            ), "Can't replace boxes in text if it's already encoded"
+
+        # Note that the input_embeds are returned in pytorch's convention (sequence first)
+        return (
+            text_attention_mask,
+            text_memory_resized,
+            inputs_embeds.transpose(0, 1),
+        )

sam3/model/tokenizer_ve.py

@@ -0,0 +1,253 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+"""
+Text Tokenizer.
+
+Copied and lightly adapted from VE repo, which in turn copied
+from open_clip and openAI CLIP.
+"""
+
+import gzip
+import html
+import io
+import os
+import string
+from functools import lru_cache
+from typing import List, Optional, Union
+
+import ftfy
+import regex as re
+import torch
+from iopath.common.file_io import g_pathmgr
+
+
+# https://stackoverflow.com/q/62691279
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+DEFAULT_CONTEXT_LENGTH = 77
+
+
+@lru_cache()
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """Return set of symbol pairs in a word.
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+def _clean_canonicalize(x):
+    # basic, remove whitespace, remove punctuation, lower case
+    return canonicalize_text(basic_clean(x))
+
+
+def _clean_lower(x):
+    # basic, remove whitespace, lower case
+    return whitespace_clean(basic_clean(x)).lower()
+
+
+def _clean_whitespace(x):
+    # basic, remove whitespace
+    return whitespace_clean(basic_clean(x))
+
+
+def get_clean_fn(type: str):
+    if type == "canonicalize":
+        return _clean_canonicalize
+    elif type == "lower":
+        return _clean_lower
+    elif type == "whitespace":
+        return _clean_whitespace
+    else:
+        assert False, f"Invalid clean function ({type})."
+
+
+def canonicalize_text(text, *, keep_punctuation_exact_string=None):
+    """Returns canonicalized `text` (lowercase and punctuation removed).
+    From: https://github.com/google-research/big_vision/blob/53f18caf27a9419231bbf08d3388b07671616d3d/big_vision/evaluators/proj/image_text/prompt_engineering.py#L94
+    Args:
+      text: string to be canonicalized.
+      keep_punctuation_exact_string: If provided, then this exact string kept.
+        For example providing '{}' will keep any occurrences of '{}' (but will
+        still remove '{' and '}' that appear separately).
+    """
+    text = text.replace("_", " ")
+    if keep_punctuation_exact_string:
+        text = keep_punctuation_exact_string.join(
+            part.translate(str.maketrans("", "", string.punctuation))
+            for part in text.split(keep_punctuation_exact_string)
+        )
+    else:
+        text = text.translate(str.maketrans("", "", string.punctuation))
+    text = text.lower()
+    text = re.sub(r"\s+", " ", text)
+    return text.strip()
+
+
+class SimpleTokenizer(object):
+    def __init__(
+        self,
+        bpe_path: Union[str, os.PathLike],
+        additional_special_tokens: Optional[List[str]] = None,
+        context_length: Optional[int] = DEFAULT_CONTEXT_LENGTH,
+        clean: str = "lower",
+    ):
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        with g_pathmgr.open(bpe_path, "rb") as fh:
+            bpe_bytes = io.BytesIO(fh.read())
+            merges = gzip.open(bpe_bytes).read().decode("utf-8").split("\n")
+        # merges = gzip.open(bpe_path).read().decode("utf-8").split("\n")
+        merges = merges[1 : 49152 - 256 - 2 + 1]
+        merges = [tuple(merge.split()) for merge in merges]
+        vocab = list(bytes_to_unicode().values())
+        vocab = vocab + [v + "</w>" for v in vocab]
+        for merge in merges:
+            vocab.append("".join(merge))
+        special_tokens = ["<start_of_text>", "<end_of_text>"]
+        if additional_special_tokens:
+            special_tokens += additional_special_tokens
+        vocab.extend(special_tokens)
+        self.encoder = dict(zip(vocab, range(len(vocab))))
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {t: t for t in special_tokens}
+        special = "|".join(special_tokens)
+        self.pat = re.compile(
+            special + r"""|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""",
+            re.IGNORECASE,
+        )
+        self.vocab_size = len(self.encoder)
+        self.all_special_ids = [self.encoder[t] for t in special_tokens]
+        self.sot_token_id = self.all_special_ids[0]
+        self.eot_token_id = self.all_special_ids[1]
+        self.context_length = context_length
+        self.clean_fn = get_clean_fn(clean)
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token[:-1]) + (token[-1] + "</w>",)
+        pairs = get_pairs(word)
+        if not pairs:
+            return token + "</w>"
+        while True:
+            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                    new_word.extend(word[i:j])
+                    i = j
+                except:
+                    new_word.extend(word[i:])
+                    break
+                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                    new_word.append(first + second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = " ".join(word)
+        self.cache[token] = word
+        return word
+
+    def encode(self, text):
+        bpe_tokens = []
+        text = self.clean_fn(text)
+        for token in re.findall(self.pat, text):
+            token = "".join(self.byte_encoder[b] for b in token.encode("utf-8"))
+            bpe_tokens.extend(
+                self.encoder[bpe_token] for bpe_token in self.bpe(token).split(" ")
+            )
+        return bpe_tokens
+
+    def decode(self, tokens):
+        text = "".join([self.decoder[token] for token in tokens])
+        text = (
+            bytearray([self.byte_decoder[c] for c in text])
+            .decode("utf-8", errors="replace")
+            .replace("</w>", " ")
+        )
+        return text
+
+    def __call__(
+        self, texts: Union[str, List[str]], context_length: Optional[int] = None
+    ) -> torch.LongTensor:
+        """Returns the tokenized representation of given input string(s)
+        Parameters
+        ----------
+        texts : Union[str, List[str]]
+            An input string or a list of input strings to tokenize
+        context_length : int
+            The context length to use; all CLIP models use 77 as the context length
+        Returns
+        -------
+        A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length]
+        """
+        if isinstance(texts, str):
+            texts = [texts]
+        context_length = context_length or self.context_length
+        assert context_length, "Please set a valid context length"
+        all_tokens = [
+            [self.sot_token_id] + self.encode(text) + [self.eot_token_id]
+            for text in texts
+        ]
+        result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)
+        for i, tokens in enumerate(all_tokens):
+            if len(tokens) > context_length:
+                tokens = tokens[:context_length]  # Truncate
+                tokens[-1] = self.eot_token_id
+            result[i, : len(tokens)] = torch.tensor(tokens)
+        return result

sam3/model/utils/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

sam3/model/utils/misc.py

@@ -0,0 +1,77 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+from collections import defaultdict
+from dataclasses import fields, is_dataclass
+from typing import Any, Mapping, Protocol, runtime_checkable
+
+import torch
+
+
+def _is_named_tuple(x) -> bool:
+    return isinstance(x, tuple) and hasattr(x, "_asdict") and hasattr(x, "_fields")
+
+
+@runtime_checkable
+class _CopyableData(Protocol):
+    def to(self, device: torch.device, *args: Any, **kwargs: Any):
+        """Copy data to the specified device"""
+        ...
+
+
+def copy_data_to_device(data, device: torch.device, *args: Any, **kwargs: Any):
+    """Function that recursively copies data to a torch.device.
+
+    Args:
+        data: The data to copy to device
+        device: The device to which the data should be copied
+        args: positional arguments that will be passed to the `to` call
+        kwargs: keyword arguments that will be passed to the `to` call
+
+    Returns:
+        The data on the correct device
+    """
+
+    if _is_named_tuple(data):
+        return type(data)(
+            **copy_data_to_device(data._asdict(), device, *args, **kwargs)
+        )
+    elif isinstance(data, (list, tuple)):
+        return type(data)(copy_data_to_device(e, device, *args, **kwargs) for e in data)
+    elif isinstance(data, defaultdict):
+        return type(data)(
+            data.default_factory,
+            {
+                k: copy_data_to_device(v, device, *args, **kwargs)
+                for k, v in data.items()
+            },
+        )
+    elif isinstance(data, Mapping):
+        return type(data)(
+            {
+                k: copy_data_to_device(v, device, *args, **kwargs)
+                for k, v in data.items()
+            }
+        )
+    elif is_dataclass(data) and not isinstance(data, type):
+        new_data_class = type(data)(
+            **{
+                field.name: copy_data_to_device(
+                    getattr(data, field.name), device, *args, **kwargs
+                )
+                for field in fields(data)
+                if field.init
+            }
+        )
+        for field in fields(data):
+            if not field.init:
+                setattr(
+                    new_data_class,
+                    field.name,
+                    copy_data_to_device(
+                        getattr(data, field.name), device, *args, **kwargs
+                    ),
+                )
+        return new_data_class
+    elif isinstance(data, _CopyableData):
+        return data.to(device, *args, **kwargs)
+    return data

sam3/model/utils/sam1_utils.py

@@ -0,0 +1,119 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import warnings
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.transforms import Normalize, Resize, ToTensor
+
+
+# Adapted from https://github.com/facebookresearch/sam2/blob/main/sam2/utils/transforms.py
+class SAM2Transforms(nn.Module):
+    def __init__(
+        self, resolution, mask_threshold, max_hole_area=0.0, max_sprinkle_area=0.0
+    ):
+        """
+        Transforms for SAM2.
+        """
+        super().__init__()
+        self.resolution = resolution
+        self.mask_threshold = mask_threshold
+        self.max_hole_area = max_hole_area
+        self.max_sprinkle_area = max_sprinkle_area
+        self.mean = [0.5, 0.5, 0.5]
+        self.std = [0.5, 0.5, 0.5]
+        self.to_tensor = ToTensor()
+        self.transforms = torch.jit.script(
+            nn.Sequential(
+                Resize((self.resolution, self.resolution)),
+                Normalize(self.mean, self.std),
+            )
+        )
+
+    def __call__(self, x):
+        x = self.to_tensor(x)
+        return self.transforms(x)
+
+    def forward_batch(self, img_list):
+        img_batch = [self.transforms(self.to_tensor(img)) for img in img_list]
+        img_batch = torch.stack(img_batch, dim=0)
+        return img_batch
+
+    def transform_coords(
+        self, coords: torch.Tensor, normalize=False, orig_hw=None
+    ) -> torch.Tensor:
+        """
+        Expects a torch tensor with length 2 in the last dimension. The coordinates can be in absolute image or normalized coordinates,
+        If the coords are in absolute image coordinates, normalize should be set to True and original image size is required.
+
+        Returns
+            Un-normalized coordinates in the range of [0, 1] which is expected by the SAM2 model.
+        """
+        if normalize:
+            assert orig_hw is not None
+            h, w = orig_hw
+            coords = coords.clone()
+            coords[..., 0] = coords[..., 0] / w
+            coords[..., 1] = coords[..., 1] / h
+
+        coords = coords * self.resolution  # unnormalize coords
+        return coords
+
+    def transform_boxes(
+        self, boxes: torch.Tensor, normalize=False, orig_hw=None
+    ) -> torch.Tensor:
+        """
+        Expects a tensor of shape Bx4. The coordinates can be in absolute image or normalized coordinates,
+        if the coords are in absolute image coordinates, normalize should be set to True and original image size is required.
+        """
+        boxes = self.transform_coords(boxes.reshape(-1, 2, 2), normalize, orig_hw)
+        return boxes
+
+    def postprocess_masks(self, masks: torch.Tensor, orig_hw) -> torch.Tensor:
+        """
+        Perform PostProcessing on output masks.
+        """
+        masks = masks.float()
+        input_masks = masks
+        mask_flat = masks.flatten(0, 1).unsqueeze(1)  # flatten as 1-channel image
+        try:
+            from sam3.perflib.connected_components import connected_components
+
+            if self.max_hole_area > 0:
+                # Holes are those connected components in background with area <= self.fill_hole_area
+                # (background regions are those with mask scores <= self.mask_threshold)
+                labels, areas = connected_components(
+                    (mask_flat <= self.mask_threshold).to(torch.uint8)
+                )
+                is_hole = (labels > 0) & (areas <= self.max_hole_area)
+                is_hole = is_hole.reshape_as(masks)
+                # We fill holes with a small positive mask score (10.0) to change them to foreground.
+                masks = torch.where(is_hole, self.mask_threshold + 10.0, masks)
+
+            if self.max_sprinkle_area > 0:
+                labels, areas = connected_components(
+                    (mask_flat > self.mask_threshold).to(torch.uint8)
+                )
+                is_hole = (labels > 0) & (areas <= self.max_sprinkle_area)
+                is_hole = is_hole.reshape_as(masks)
+                # We fill holes with negative mask score (-10.0) to change them to background.
+                masks = torch.where(is_hole, self.mask_threshold - 10.0, masks)
+        except Exception as e:
+            # Skip the post-processing step if the CUDA kernel fails
+            warnings.warn(
+                f"{e}\n\nSkipping the post-processing step due to the error above. You can "
+                "still use SAM 3 and it's OK to ignore the error above, although some post-processing "
+                "functionality may be limited (which doesn't affect the results in most cases; see "
+                "https://github.com/facebookresearch/sam3/blob/main/INSTALL.md).",
+                category=UserWarning,
+                stacklevel=2,
+            )
+            masks = input_masks
+
+        masks = F.interpolate(masks, orig_hw, mode="bilinear", align_corners=False)
+        return masks

sam3/model/utils/sam2_utils.py

@@ -0,0 +1,233 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+from threading import Thread
+
+import numpy as np
+import torch
+from PIL import Image
+from tqdm import tqdm
+
+
+def _load_img_as_tensor(img_path, image_size):
+    img_pil = Image.open(img_path)
+    img_np = np.array(img_pil.convert("RGB").resize((image_size, image_size)))
+    if img_np.dtype == np.uint8:  # np.uint8 is expected for JPEG images
+        img_np = img_np / 255.0
+    else:
+        raise RuntimeError(f"Unknown image dtype: {img_np.dtype} on {img_path}")
+    img = torch.from_numpy(img_np).permute(2, 0, 1)
+    video_width, video_height = img_pil.size  # the original video size
+    return img, video_height, video_width
+
+
+class AsyncVideoFrameLoader:
+    """
+    A list of video frames to be load asynchronously without blocking session start.
+    """
+
+    def __init__(
+        self,
+        img_paths,
+        image_size,
+        offload_video_to_cpu,
+        img_mean,
+        img_std,
+        compute_device,
+    ):
+        self.img_paths = img_paths
+        self.image_size = image_size
+        self.offload_video_to_cpu = offload_video_to_cpu
+        self.img_mean = img_mean
+        self.img_std = img_std
+        # items in `self.images` will be loaded asynchronously
+        self.images = [None] * len(img_paths)
+        # catch and raise any exceptions in the async loading thread
+        self.exception = None
+        # video_height and video_width be filled when loading the first image
+        self.video_height = None
+        self.video_width = None
+        self.compute_device = compute_device
+
+        # load the first frame to fill video_height and video_width and also
+        # to cache it (since it's most likely where the user will click)
+        self.__getitem__(0)
+
+        # load the rest of frames asynchronously without blocking the session start
+        def _load_frames():
+            try:
+                for n in tqdm(range(len(self.images)), desc="frame loading (JPEG)"):
+                    self.__getitem__(n)
+            except Exception as e:
+                self.exception = e
+
+        self.thread = Thread(target=_load_frames, daemon=True)
+        self.thread.start()
+
+    def __getitem__(self, index):
+        if self.exception is not None:
+            raise RuntimeError("Failure in frame loading thread") from self.exception
+
+        img = self.images[index]
+        if img is not None:
+            return img
+
+        img, video_height, video_width = _load_img_as_tensor(
+            self.img_paths[index], self.image_size
+        )
+        self.video_height = video_height
+        self.video_width = video_width
+        # normalize by mean and std
+        img -= self.img_mean
+        img /= self.img_std
+        if not self.offload_video_to_cpu:
+            img = img.to(self.compute_device, non_blocking=True)
+        self.images[index] = img
+        return img
+
+    def __len__(self):
+        return len(self.images)
+
+
+def load_video_frames(
+    video_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.485, 0.456, 0.406),
+    img_std=(0.229, 0.224, 0.225),
+    async_loading_frames=False,
+    compute_device=torch.device("cuda"),
+):
+    """
+    Load the video frames from video_path. The frames are resized to image_size as in
+    the model and are loaded to GPU if offload_video_to_cpu=False. This is used by the demo.
+    """
+    is_bytes = isinstance(video_path, bytes)
+    is_str = isinstance(video_path, str)
+    is_mp4_path = is_str and os.path.splitext(video_path)[-1] in [".mp4", ".MP4"]
+    if is_bytes or is_mp4_path:
+        return load_video_frames_from_video_file(
+            video_path=video_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+            compute_device=compute_device,
+        )
+    elif is_str and os.path.isdir(video_path):
+        return load_video_frames_from_jpg_images(
+            video_path=video_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+            async_loading_frames=async_loading_frames,
+            compute_device=compute_device,
+        )
+    else:
+        raise NotImplementedError(
+            "Only MP4 video and JPEG folder are supported at this moment"
+        )
+
+
+def load_video_frames_from_jpg_images(
+    video_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.485, 0.456, 0.406),
+    img_std=(0.229, 0.224, 0.225),
+    async_loading_frames=False,
+    compute_device=torch.device("cuda"),
+):
+    """
+    Load the video frames from a directory of JPEG files ("<frame_index>.jpg" format).
+
+    The frames are resized to image_size x image_size and are loaded to GPU if
+    `offload_video_to_cpu` is `False` and to CPU if `offload_video_to_cpu` is `True`.
+
+    You can load a frame asynchronously by setting `async_loading_frames` to `True`.
+    """
+    if isinstance(video_path, str) and os.path.isdir(video_path):
+        jpg_folder = video_path
+    else:
+        raise NotImplementedError(
+            "Only JPEG frames are supported at this moment. For video files, you may use "
+            "ffmpeg (https://ffmpeg.org/) to extract frames into a folder of JPEG files, such as \n"
+            "```\n"
+            "ffmpeg -i <your_video>.mp4 -q:v 2 -start_number 0 <output_dir>/'%05d.jpg'\n"
+            "```\n"
+            "where `-q:v` generates high-quality JPEG frames and `-start_number 0` asks "
+            "ffmpeg to start the JPEG file from 00000.jpg."
+        )
+
+    frame_names = [
+        p
+        for p in os.listdir(jpg_folder)
+        if os.path.splitext(p)[-1] in [".jpg", ".jpeg", ".JPG", ".JPEG"]
+    ]
+    frame_names.sort(key=lambda p: int(os.path.splitext(p)[0]))
+    num_frames = len(frame_names)
+    if num_frames == 0:
+        raise RuntimeError(f"no images found in {jpg_folder}")
+    img_paths = [os.path.join(jpg_folder, frame_name) for frame_name in frame_names]
+    img_mean = torch.tensor(img_mean, dtype=torch.float32)[:, None, None]
+    img_std = torch.tensor(img_std, dtype=torch.float32)[:, None, None]
+
+    if async_loading_frames:
+        lazy_images = AsyncVideoFrameLoader(
+            img_paths,
+            image_size,
+            offload_video_to_cpu,
+            img_mean,
+            img_std,
+            compute_device,
+        )
+        return lazy_images, lazy_images.video_height, lazy_images.video_width
+
+    images = torch.zeros(num_frames, 3, image_size, image_size, dtype=torch.float32)
+    for n, img_path in enumerate(tqdm(img_paths, desc="frame loading (JPEG)")):
+        images[n], video_height, video_width = _load_img_as_tensor(img_path, image_size)
+    if not offload_video_to_cpu:
+        images = images.to(compute_device)
+        img_mean = img_mean.to(compute_device)
+        img_std = img_std.to(compute_device)
+    # normalize by mean and std
+    images -= img_mean
+    images /= img_std
+    return images, video_height, video_width
+
+
+def load_video_frames_from_video_file(
+    video_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.485, 0.456, 0.406),
+    img_std=(0.229, 0.224, 0.225),
+    compute_device=torch.device("cuda"),
+):
+    """Load the video frames from a video file."""
+    import decord
+
+    img_mean = torch.tensor(img_mean, dtype=torch.float32)[:, None, None]
+    img_std = torch.tensor(img_std, dtype=torch.float32)[:, None, None]
+    # Get the original video height and width
+    decord.bridge.set_bridge("torch")
+    video_height, video_width, _ = decord.VideoReader(video_path).next().shape
+    # Iterate over all frames in the video
+    images = []
+    for frame in decord.VideoReader(video_path, width=image_size, height=image_size):
+        images.append(frame.permute(2, 0, 1))
+
+    images = torch.stack(images, dim=0).float() / 255.0
+    if not offload_video_to_cpu:
+        images = images.to(compute_device)
+        img_mean = img_mean.to(compute_device)
+        img_std = img_std.to(compute_device)
+    # normalize by mean and std
+    images -= img_mean
+    images /= img_std
+    return images, video_height, video_width

sam3/model/vitdet.py

@@ -0,0 +1,879 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+"""
+ViTDet backbone adapted from Detectron2.
+This module implements Vision Transformer (ViT) backbone for object detection.
+
+Rope embedding code adopted from:
+1. https://github.com/meta-llama/codellama/blob/main/llama/model.py
+2. https://github.com/naver-ai/rope-vit
+3. https://github.com/lucidrains/rotary-embedding-torch
+"""
+
+import math
+from functools import partial
+from typing import Callable, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+
+try:
+    from timm.layers import DropPath, Mlp, trunc_normal_
+except ModuleNotFoundError:
+    # compatibility for older timm versions
+    from timm.models.layers import DropPath, Mlp, trunc_normal_
+from torch import Tensor
+
+from .model_misc import LayerScale
+
+
+def init_t_xy(
+    end_x: int, end_y: int, scale: float = 1.0, offset: int = 0
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    t = torch.arange(end_x * end_y, dtype=torch.float32)
+    t_x = (t % end_x).float()
+    t_y = torch.div(t, end_x, rounding_mode="floor").float()
+    return t_x * scale + offset, t_y * scale + offset
+
+
+def compute_axial_cis(
+    dim: int,
+    end_x: int,
+    end_y: int,
+    theta: float = 10000.0,
+    scale_pos: float = 1.0,
+    offset: int = 0,
+) -> torch.Tensor:
+    freqs_x = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
+    freqs_y = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
+
+    t_x, t_y = init_t_xy(end_x, end_y, scale_pos, offset)
+    freqs_x = torch.outer(t_x, freqs_x)
+    freqs_y = torch.outer(t_y, freqs_y)
+    freqs_cis_x = torch.polar(torch.ones_like(freqs_x), freqs_x)
+    freqs_cis_y = torch.polar(torch.ones_like(freqs_y), freqs_y)
+    return torch.cat([freqs_cis_x, freqs_cis_y], dim=-1)
+
+
+def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor) -> torch.Tensor:
+    ndim = x.ndim
+    assert 0 <= 1 < ndim
+    assert freqs_cis.shape == (x.shape[-2], x.shape[-1])
+    shape = [d if i >= ndim - 2 else 1 for i, d in enumerate(x.shape)]
+    return freqs_cis.view(*shape)
+
+
+def apply_rotary_enc(
+    xq: torch.Tensor,
+    xk: torch.Tensor,
+    freqs_cis: torch.Tensor,
+    repeat_freqs_k: bool = False,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
+    xk_ = (
+        torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
+        if xk.shape[-2] != 0
+        else None
+    )
+    freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
+    xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
+    if xk_ is None:
+        # no keys to rotate, due to dropout
+        return xq_out.type_as(xq).to(xq.device), xk
+    # repeat freqs along seq_len dim to match k seq_len
+    if repeat_freqs_k:
+        r = xk_.shape[-2] // xq_.shape[-2]
+        freqs_cis = freqs_cis.repeat(*([1] * (freqs_cis.ndim - 2)), r, 1)
+    xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
+    return xq_out.type_as(xq).to(xq.device), xk_out.type_as(xk).to(xk.device)
+
+
+def window_partition(x: Tensor, window_size: int) -> Tuple[Tensor, Tuple[int, int]]:
+    """
+    Partition into non-overlapping windows with padding if needed.
+    Args:
+        x (tensor): input tokens with [B, H, W, C].
+        window_size (int): window size.
+    Returns:
+        windows: windows after partition with [B * num_windows, window_size, window_size, C].
+        (Hp, Wp): padded height and width before partition
+    """
+    B, H, W, C = x.shape
+
+    pad_h = (window_size - H % window_size) % window_size
+    pad_w = (window_size - W % window_size) % window_size
+    if pad_h > 0 or pad_w > 0:
+        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
+    Hp, Wp = H + pad_h, W + pad_w
+
+    x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).reshape(-1, window_size, window_size, C)
+    return windows, (Hp, Wp)
+
+
+def window_unpartition(
+    windows: Tensor, window_size: int, pad_hw: Tuple[int, int], hw: Tuple[int, int]
+) -> Tensor:
+    """
+    Window unpartition into original sequences and removing padding.
+    Args:
+        x (tensor): input tokens with [B * num_windows, window_size, window_size, C].
+        window_size (int): window size.
+        pad_hw (Tuple): padded height and width (Hp, Wp).
+        hw (Tuple): original height and width (H, W) before padding.
+    Returns:
+        x: unpartitioned sequences with [B, H, W, C].
+    """
+    Hp, Wp = pad_hw
+    H, W = hw
+    B = windows.shape[0] // (Hp * Wp // window_size // window_size)
+    x = windows.reshape(
+        B, Hp // window_size, Wp // window_size, window_size, window_size, -1
+    )
+    x = x.permute(0, 1, 3, 2, 4, 5).reshape(B, Hp, Wp, -1)
+
+    if Hp > H or Wp > W:
+        x = x[:, :H, :W, :]
+    return x
+
+
+def get_rel_pos(q_size: int, k_size: int, rel_pos: Tensor) -> Tensor:
+    """
+    Get relative positional embeddings according to the relative positions of
+        query and key sizes.
+    Args:
+        q_size (int): size of query q.
+        k_size (int): size of key k.
+        rel_pos (Tensor): relative position embeddings (L, C).
+    Returns:
+        Extracted positional embeddings according to relative positions.
+    """
+    max_rel_dist = int(2 * max(q_size, k_size) - 1)
+    # Interpolate rel pos if needed.
+    if rel_pos.shape[0] != max_rel_dist:
+        # Interpolate rel pos.
+        rel_pos_resized = F.interpolate(
+            rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
+            size=max_rel_dist,
+            mode="linear",
+            align_corners=False,
+        )
+        rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
+    else:
+        rel_pos_resized = rel_pos
+
+    # Scale the coords with short length if shapes for q and k are different.
+    q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
+    k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
+    relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
+
+    return rel_pos_resized[relative_coords.long()]
+
+
+def get_abs_pos(
+    abs_pos: Tensor,
+    has_cls_token: bool,
+    hw: Tuple[int, int],
+    retain_cls_token: bool = False,
+    tiling: bool = False,
+) -> Tensor:
+    """
+    Calculate absolute positional embeddings. If needed, resize embeddings and remove cls_token
+        dimension for the original embeddings.
+    Args:
+        abs_pos (Tensor): absolute positional embeddings with (1, num_position, C).
+        has_cls_token (bool): If true, has 1 embedding in abs_pos for cls token.
+        hw (Tuple): size of input image tokens.
+        retain_cls_token: whether to retain the cls_token
+        tiling: whether to tile the embeddings, *instead* of interpolation (a la abs_win)
+    Returns:
+        Absolute positional embeddings after processing with shape (1, H, W, C),
+        if retain_cls_token is False, otherwise (1, 1+H*W, C)
+    """
+    if retain_cls_token:
+        assert has_cls_token
+
+    h, w = hw
+    if has_cls_token:
+        cls_pos = abs_pos[:, :1]
+        abs_pos = abs_pos[:, 1:]
+
+    xy_num = abs_pos.shape[1]
+    size = int(math.sqrt(xy_num))
+    assert size * size == xy_num
+
+    if size != h or size != w:
+        new_abs_pos = abs_pos.reshape(1, size, size, -1).permute(0, 3, 1, 2)
+        if tiling:
+            new_abs_pos = new_abs_pos.tile(
+                [1, 1] + [x // y + 1 for x, y in zip((h, w), new_abs_pos.shape[2:])]
+            )[:, :, :h, :w]
+        else:
+            new_abs_pos = F.interpolate(
+                new_abs_pos,
+                size=(h, w),
+                mode="bicubic",
+                align_corners=False,
+            )
+
+        if not retain_cls_token:
+            return new_abs_pos.permute(0, 2, 3, 1)
+        else:
+            # add cls_token back, flatten spatial dims
+            assert has_cls_token
+            return torch.cat(
+                [cls_pos, new_abs_pos.permute(0, 2, 3, 1).reshape(1, h * w, -1)],
+                dim=1,
+            )
+
+    else:
+        if not retain_cls_token:
+            return abs_pos.reshape(1, h, w, -1)
+        else:
+            assert has_cls_token
+            return torch.cat([cls_pos, abs_pos], dim=1)
+
+
+def concat_rel_pos(
+    q: Tensor,
+    k: Tensor,
+    q_hw: Tuple[int, int],
+    k_hw: Tuple[int, int],
+    rel_pos_h: Tensor,
+    rel_pos_w: Tensor,
+    rescale: bool = False,
+    relative_coords: Optional[Tensor] = None,
+) -> Tuple[Tensor, Tensor]:
+    """
+    Concatenate rel pos coeffs to the q & k tensors, so that qk^T is now
+    effectively including rel pos biases.
+    Args:
+        q (Tensor): q tensor with shape (B, L_q, C).
+        k (Tensor): k tensor with shape (B, L_k, C).
+        q_hw, k_hw: These are spatial size of q & k tensors.
+        rel_pos_h, rel_pos_w: These are relative pos embeddings/params of height, width.
+        rescale (bool): whether to rescale. e.g. for use when using sdpa, pytorch will
+            scale by the wrong factor due to the concat.
+    Returns:
+        q, k: But, padded so that qk^T accounts for rel pos biases
+    """
+    q_h, q_w = q_hw
+    k_h, k_w = k_hw
+
+    assert (q_h == q_w) and (k_h == k_w), "only square inputs supported"
+
+    if relative_coords is not None:
+        Rh = rel_pos_h[relative_coords]
+        Rw = rel_pos_w[relative_coords]
+    else:
+        Rh = get_rel_pos(q_h, k_h, rel_pos_h)
+        Rw = get_rel_pos(q_w, k_w, rel_pos_w)
+
+    B, _, dim = q.shape
+    r_q = q.reshape(B, q_h, q_w, dim)
+
+    old_scale = dim**0.5
+    new_scale = (dim + k_h + k_w) ** 0.5 if rescale else old_scale  # for sdpa
+    # attn will be divided by new_scale, but we want to divide q by old_scale
+    scale_ratio = new_scale / old_scale
+
+    rel_h = torch.einsum("bhwc,hkc->bhwk", r_q, Rh) * new_scale  # (B, q_h, q_w, k_h)
+    rel_w = torch.einsum("bhwc,wkc->bhwk", r_q, Rw) * new_scale  # (B, q_h, q_w, k_w)
+
+    eye_h = torch.eye(k_h, dtype=q.dtype, device=q.device)
+    eye_w = torch.eye(k_w, dtype=q.dtype, device=q.device)
+
+    eye_h = eye_h.view(1, k_h, 1, k_h).expand([B, k_h, k_w, k_h])
+    eye_w = eye_w.view(1, 1, k_w, k_w).expand([B, k_h, k_w, k_w])
+
+    q = torch.cat([r_q * scale_ratio, rel_h, rel_w], dim=-1).view(B, q_h * q_w, -1)
+    k = torch.cat([k.view(B, k_h, k_w, -1), eye_h, eye_w], dim=-1).view(
+        B, k_h * k_w, -1
+    )
+
+    return q, k
+
+
+class PatchEmbed(nn.Module):
+    """
+    Image to Patch Embedding.
+    """
+
+    def __init__(
+        self,
+        kernel_size: Tuple[int, int] = (16, 16),
+        stride: Tuple[int, int] = (16, 16),
+        padding: Tuple[int, int] = (0, 0),
+        in_chans: int = 3,
+        embed_dim: int = 768,
+        bias: bool = True,
+    ):
+        """
+        Args:
+            kernel_size (Tuple): kernel size of the projection layer.
+            stride (Tuple): stride of the projection layer.
+            padding (Tuple): padding size of the projection layer.
+            in_chans (int): Number of input image channels.
+            embed_dim (int):  embed_dim (int): Patch embedding dimension.
+        """
+        super().__init__()
+
+        self.proj = nn.Conv2d(
+            in_chans,
+            embed_dim,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            bias=bias,
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = self.proj(x)
+        # B C H W -> B H W C
+        x = x.permute(0, 2, 3, 1)
+        return x
+
+
+class Attention(nn.Module):
+    """Multi-head Attention block with relative position embeddings and 2d-rope."""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        qkv_bias: bool = True,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        input_size: Optional[Tuple[int, int]] = None,
+        cls_token: bool = False,
+        use_rope: bool = False,
+        rope_theta: float = 10000.0,
+        rope_pt_size: Optional[Tuple[int, int]] = None,
+        rope_interp: bool = False,
+    ):
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads.
+            qkv_bias (bool:  If True, add a learnable bias to query, key, value.
+            rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            input_size (int or None): Input resolution for calculating the relative positional
+                parameter size or rope size.
+            attn_type: Type of attention operation, e.g. "vanilla", "vanilla-xformer".
+            cls_token: whether a cls_token is present.
+            use_rope: whether to use rope 2d (indep of use_rel_pos, as it can be used together)
+            rope_theta: control frequencies of rope
+            rope_pt_size: size of rope in previous stage of training, needed for interpolation or tiling
+            rope_interp: whether to interpolate (or extrapolate) rope to match input size
+        """
+        super().__init__()
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.scale = self.head_dim**-0.5
+        self.cls_token = cls_token
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.proj = nn.Linear(dim, dim)
+
+        # rel_pos embeddings and rope
+        self.use_rel_pos = use_rel_pos
+        self.input_size = input_size
+
+        self.use_rope = use_rope
+        self.rope_theta = rope_theta
+        self.rope_pt_size = rope_pt_size
+        self.rope_interp = rope_interp
+
+        # init rel_pos embeddings and rope
+        self._setup_rel_pos(rel_pos_zero_init)
+        self._setup_rope_freqs()
+
+    def _setup_rel_pos(self, rel_pos_zero_init: bool = True) -> None:
+        if not self.use_rel_pos:
+            self.rel_pos_h = None
+            self.rel_pos_w = None
+            return
+
+        assert self.input_size is not None
+        assert self.cls_token is False, "not supported"
+        # initialize relative positional embeddings
+        self.rel_pos_h = nn.Parameter(
+            torch.zeros(2 * self.input_size[0] - 1, self.head_dim)
+        )
+        self.rel_pos_w = nn.Parameter(
+            torch.zeros(2 * self.input_size[1] - 1, self.head_dim)
+        )
+
+        if not rel_pos_zero_init:
+            trunc_normal_(self.rel_pos_h, std=0.02)
+            trunc_normal_(self.rel_pos_w, std=0.02)
+
+        # Precompute the relative coords
+        H, W = self.input_size
+        q_coords = torch.arange(H)[:, None]
+        k_coords = torch.arange(W)[None, :]
+        relative_coords = (q_coords - k_coords) + (H - 1)
+        self.register_buffer("relative_coords", relative_coords.long())
+
+    def _setup_rope_freqs(self) -> None:
+        if not self.use_rope:
+            self.freqs_cis = None
+            return
+
+        assert self.input_size is not None
+        # determine rope input size
+        if self.rope_pt_size is None:
+            self.rope_pt_size = self.input_size
+
+        # initialize 2d rope freqs
+        self.compute_cis = partial(
+            compute_axial_cis,
+            dim=self.head_dim,
+            theta=self.rope_theta,
+        )
+
+        # interpolate rope
+        scale_pos = 1.0
+        if self.rope_interp:
+            scale_pos = self.rope_pt_size[0] / self.input_size[0]
+        # get scaled freqs_cis
+        freqs_cis = self.compute_cis(
+            end_x=self.input_size[0],
+            end_y=self.input_size[1],
+            scale_pos=scale_pos,
+        )
+        if self.cls_token:
+            t = torch.zeros(
+                self.head_dim // 2,
+                dtype=torch.float32,
+                device=freqs_cis.device,
+            )
+            cls_freqs_cis = torch.polar(torch.ones_like(t), t)[None, :]
+            freqs_cis = torch.cat([cls_freqs_cis, freqs_cis], dim=0)
+
+        self.register_buffer("freqs_cis", freqs_cis)
+
+    def _apply_rope(self, q, k) -> Tuple[Tensor, Tensor]:
+        if not self.use_rope:
+            return q, k
+
+        assert self.freqs_cis is not None
+        return apply_rotary_enc(q, k, freqs_cis=self.freqs_cis)
+
+    def forward(self, x: Tensor) -> Tensor:
+        s = 1 if self.cls_token else 0  # used to exclude cls_token
+        if x.ndim == 4:
+            B, H, W, _ = x.shape
+            assert s == 0  # no cls_token
+            L = H * W
+            ndim = 4
+        else:
+            assert x.ndim == 3
+            B, L, _ = x.shape
+            ndim = 3
+            H = W = math.sqrt(L - s)
+
+        # qkv with shape (3, B, nHead, L, C)
+        qkv = self.qkv(x).reshape(B, L, 3, self.num_heads, -1)
+        # q, k, v with shape (B, nHead, L, C)
+        q, k, v = qkv.permute(2, 0, 3, 1, 4).unbind(0)
+
+        # handle rope and rel pos embeddings
+        q, k = self._apply_rope(q, k)
+        if self.use_rel_pos:
+            q, k = concat_rel_pos(
+                q.flatten(0, 1),
+                k.flatten(0, 1),
+                (H, W),
+                x.shape[1:3],
+                self.rel_pos_h,
+                self.rel_pos_w,
+                rescale=True,
+                relative_coords=self.relative_coords,
+            )
+
+            # sdpa expects [B, nheads, H*W, C] so we transpose back
+            q = q.reshape(B, self.num_heads, H * W, -1)
+            k = k.reshape(B, self.num_heads, H * W, -1)
+
+        x = F.scaled_dot_product_attention(q, k, v)
+
+        if ndim == 4:
+            x = (
+                x.view(B, self.num_heads, H, W, -1)
+                .permute(0, 2, 3, 1, 4)
+                .reshape(B, H, W, -1)
+            )
+        else:
+            x = x.view(B, self.num_heads, L, -1).permute(0, 2, 1, 3).reshape(B, L, -1)
+
+        x = self.proj(x)
+
+        return x
+
+
+class Block(nn.Module):
+    """Transformer blocks with support of window attention"""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        drop_path: float = 0.0,
+        norm_layer: Callable[..., nn.Module] = nn.LayerNorm,
+        act_layer: Callable[..., nn.Module] = nn.GELU,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        input_size: Optional[Tuple[int, int]] = None,
+        use_rope: bool = False,
+        rope_pt_size: Optional[Tuple[int, int]] = None,
+        rope_tiled: bool = False,
+        rope_interp: bool = False,
+        use_ve_rope: bool = False,
+        cls_token: bool = False,
+        dropout: float = 0.0,
+        init_values: Optional[float] = None,
+    ):
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            drop_path (float): Stochastic depth rate.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks. If it equals 0, then not
+                use window attention.
+            input_size (int or None): Input resolution for calculating the relative positional
+                parameter size.
+            dropout (float): Dropout rate.
+            cls_token: whether a cls_token is present.
+            use_rope: whether to use rope 2d (indep of use_rel_pos, as it can be used together)
+            rope_pt_size: size of rope in previous stage of training, needed for interpolation or tiling
+            rope_interp: whether to interpolate (or extrapolate) rope to match target input size,
+                expected to specify source size as rope_pt_size.
+        """
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            use_rel_pos=use_rel_pos,
+            rel_pos_zero_init=rel_pos_zero_init,
+            input_size=input_size if window_size == 0 else (window_size, window_size),
+            use_rope=use_rope,
+            rope_pt_size=rope_pt_size,
+            rope_interp=rope_interp,
+            cls_token=cls_token,
+        )
+        self.ls1 = (
+            LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
+        )
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+        self.norm2 = norm_layer(dim)
+        self.mlp = Mlp(
+            in_features=dim,
+            hidden_features=int(dim * mlp_ratio),
+            act_layer=act_layer,
+            drop=(dropout, 0.0),
+        )
+        self.ls2 = (
+            LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
+        )
+        self.dropout = nn.Dropout(dropout)
+        self.window_size = window_size
+
+    def forward(self, x: Tensor) -> Tensor:
+        shortcut = x
+        x = self.norm1(x)
+        # Window partition
+        if self.window_size > 0:
+            H, W = x.shape[1], x.shape[2]
+            x, pad_hw = window_partition(x, self.window_size)
+
+        x = self.ls1(self.attn(x))
+        # Reverse window partition
+        if self.window_size > 0:
+            x = window_unpartition(x, self.window_size, pad_hw, (H, W))
+
+        x = shortcut + self.dropout(self.drop_path(x))
+        x = x + self.dropout(self.drop_path(self.ls2(self.mlp(self.norm2(x)))))
+
+        return x
+
+
+class ViT(nn.Module):
+    """
+    This module implements Vision Transformer (ViT) backbone in :paper:`vitdet`.
+    "Exploring Plain Vision Transformer Backbones for Object Detection",
+    https://arxiv.org/abs/2203.16527
+    """
+
+    def __init__(
+        self,
+        img_size: int = 1024,
+        patch_size: int = 16,
+        in_chans: int = 3,
+        embed_dim: int = 768,
+        depth: int = 12,
+        num_heads: int = 12,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        drop_path_rate: float = 0.0,
+        norm_layer: Union[Callable[..., nn.Module], str] = "LayerNorm",
+        act_layer: Callable[..., nn.Module] = nn.GELU,
+        use_abs_pos: bool = True,
+        tile_abs_pos: bool = True,
+        rel_pos_blocks: Union[Tuple[int, ...], bool] = (2, 5, 8, 11),
+        rel_pos_zero_init: bool = True,
+        window_size: int = 14,
+        global_att_blocks: Tuple[int, ...] = (2, 5, 8, 11),
+        use_rope: bool = False,
+        rope_pt_size: Optional[int] = None,
+        use_interp_rope: bool = False,
+        pretrain_img_size: int = 224,
+        pretrain_use_cls_token: bool = True,
+        retain_cls_token: bool = True,
+        dropout: float = 0.0,
+        return_interm_layers: bool = False,
+        init_values: Optional[float] = None,  # for layerscale
+        ln_pre: bool = False,
+        ln_post: bool = False,
+        bias_patch_embed: bool = True,
+        compile_mode: Optional[str] = None,
+        use_act_checkpoint: bool = True,
+    ):
+        """
+        Args:
+            img_size (int): Input image size. Only relevant for rel pos or rope.
+            patch_size (int): Patch size.
+            in_chans (int): Number of input image channels.
+            embed_dim (int): Patch embedding dimension.
+            depth (int): Depth of ViT.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            drop_path_rate (float): Stochastic depth rate.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_abs_pos (bool): If True, use absolute positional embeddings.
+            tile_abs_pos (bool): If True, tile absolute positional embeddings instead of interpolation.
+            rel_pos_blocks (list): Blocks which have rel pos embeddings.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks.
+            global_att_blocks (list): Indexes for blocks using global attention (other blocks use window attention).
+            use_rope (bool): whether to use rope 2d (indep of rel_pos_blocks, as it can be used together).
+            rope_pt_size (int): size of rope in previous stage of training, needed for interpolation or tiling.
+            use_interp_rope: whether to interpolate (or extrapolate) rope to match target input size,
+                expected to specify source size as rope_pt_size.
+            use_act_checkpoint (bool): If True, use activation checkpointing.
+            pretrain_img_size (int): input image size for pretraining models.
+            pretrain_use_cls_token (bool): If True, pretraining models use class token.
+            retain_cls_token: whether cls_token should be retained.
+            dropout (float): Dropout rate. Applied in residual blocks of attn, mlp and inside the mlp.
+
+            return_interm_layers (bool): Whether to return intermediate layers (all global attention blocks).
+            init_values: layer scale init, None for no layer scale.
+
+            ln_pre (bool): If True, apply layer norm before transformer blocks.
+            ln_post (bool): If True, apply layer norm after transformer blocks.
+            bias_patch_embed (bool): bias in conv for patch embed?
+            compile_mode (str): mode to compile the forward
+        """
+        super().__init__()
+        self.pretrain_use_cls_token = pretrain_use_cls_token
+
+        window_block_indexes = [i for i in range(depth) if i not in global_att_blocks]
+        self.full_attn_ids = list(global_att_blocks)
+        self.rel_pos_blocks = [False] * depth
+        if isinstance(rel_pos_blocks, bool) and rel_pos_blocks:
+            self.rel_pos_blocks = [True] * depth
+        else:
+            for i in rel_pos_blocks:
+                self.rel_pos_blocks[i] = True
+
+        self.retain_cls_token = retain_cls_token
+        if self.retain_cls_token:
+            assert pretrain_use_cls_token
+            assert (
+                len(window_block_indexes) == 0
+            ), "windowing not supported with cls token"
+
+            assert sum(self.rel_pos_blocks) == 0, "rel pos not supported with cls token"
+
+            scale = embed_dim**-0.5
+            self.class_embedding = nn.Parameter(scale * torch.randn(1, 1, embed_dim))
+
+        if isinstance(norm_layer, str):
+            norm_layer = partial(getattr(nn, norm_layer), eps=1e-5)
+
+        self.patch_embed = PatchEmbed(
+            kernel_size=(patch_size, patch_size),
+            stride=(patch_size, patch_size),
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+            bias=bias_patch_embed,
+        )
+
+        # Handle absolute positional embedding
+        self.tile_abs_pos = tile_abs_pos
+        self.use_abs_pos = use_abs_pos
+        if self.tile_abs_pos:
+            assert self.use_abs_pos
+
+        if self.use_abs_pos:
+            # Initialize absolute positional embedding with pretrain image size.
+            num_patches = (pretrain_img_size // patch_size) * (
+                pretrain_img_size // patch_size
+            )
+            num_positions = (num_patches + 1) if pretrain_use_cls_token else num_patches
+            self.pos_embed = nn.Parameter(torch.zeros(1, num_positions, embed_dim))
+        else:
+            self.pos_embed = None
+
+        # stochastic depth decay rule
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]
+
+        self.blocks = nn.ModuleList()
+        cur_stage = 1
+        for i in range(depth):
+            block = Block(
+                dim=embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                drop_path=dpr[i],
+                norm_layer=norm_layer,
+                act_layer=act_layer,
+                use_rel_pos=self.rel_pos_blocks[i],
+                rel_pos_zero_init=rel_pos_zero_init,
+                window_size=window_size if i in window_block_indexes else 0,
+                input_size=(img_size // patch_size, img_size // patch_size),
+                use_rope=use_rope,
+                rope_pt_size=(
+                    (window_size, window_size)
+                    if rope_pt_size is None
+                    else (rope_pt_size, rope_pt_size)
+                ),
+                rope_interp=use_interp_rope,
+                cls_token=self.retain_cls_token,
+                dropout=dropout,
+                init_values=init_values,
+            )
+
+            if i not in window_block_indexes:
+                cur_stage += 1
+
+            self.use_act_checkpoint = use_act_checkpoint
+
+            self.blocks.append(block)
+
+        self.return_interm_layers = return_interm_layers
+        self.channel_list = (
+            [embed_dim] * len(self.full_attn_ids)
+            if return_interm_layers
+            else [embed_dim]
+        )
+
+        if self.pos_embed is not None:
+            trunc_normal_(self.pos_embed, std=0.02)
+
+        self.ln_pre = norm_layer(embed_dim) if ln_pre else nn.Identity()
+        self.ln_post = norm_layer(embed_dim) if ln_post else nn.Identity()
+
+        self.apply(self._init_weights)
+
+        if compile_mode is not None:
+            self.forward = torch.compile(
+                self.forward, mode=compile_mode, fullgraph=True
+            )
+            if self.use_act_checkpoint and self.training:
+                torch._dynamo.config.optimize_ddp = False
+
+    def _init_weights(self, m: nn.Module) -> None:
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=0.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
+        x = self.patch_embed(x)
+        h, w = x.shape[1], x.shape[2]
+
+        s = 0
+        if self.retain_cls_token:
+            # If cls_token is retained, we don't
+            # maintain spatial shape
+            x = torch.cat([self.class_embedding, x.flatten(1, 2)], dim=1)
+            s = 1
+
+        if self.pos_embed is not None:
+            x = x + get_abs_pos(
+                self.pos_embed,
+                self.pretrain_use_cls_token,
+                (h, w),
+                self.retain_cls_token,
+                tiling=self.tile_abs_pos,
+            )
+
+        x = self.ln_pre(x)
+
+        outputs = []
+        for i, blk in enumerate(self.blocks):
+            if self.use_act_checkpoint and self.training:
+                x = checkpoint.checkpoint(blk, x, use_reentrant=False)
+            else:
+                x = blk(x)
+            if (i == self.full_attn_ids[-1]) or (
+                self.return_interm_layers and i in self.full_attn_ids
+            ):
+                if i == self.full_attn_ids[-1]:
+                    x = self.ln_post(x)
+
+                feats = x[:, s:]
+                if feats.ndim == 4:
+                    feats = feats.permute(0, 3, 1, 2)
+                else:
+                    assert feats.ndim == 3
+                    h = w = math.sqrt(feats.shape[1])
+                    feats = feats.reshape(
+                        feats.shape[0], h, w, feats.shape[-1]
+                    ).permute(0, 3, 1, 2)
+
+                outputs.append(feats)
+
+        return outputs
+
+    def get_layer_id(self, layer_name: str) -> int:
+        # https://github.com/microsoft/unilm/blob/master/beit/optim_factory.py#L33
+        num_layers = self.get_num_layers()
+
+        if layer_name.find("rel_pos") != -1:
+            return num_layers + 1
+        elif layer_name.find("ln_pre") != -1:
+            return 0
+        elif layer_name.find("pos_embed") != -1 or layer_name.find("cls_token") != -1:
+            return 0
+        elif layer_name.find("patch_embed") != -1:
+            return 0
+        elif layer_name.find("blocks") != -1:
+            return int(layer_name.split("blocks")[1].split(".")[1]) + 1
+        else:
+            return num_layers + 1
+
+    def get_num_layers(self) -> int:
+        return len(self.blocks)

sam3/model/vl_combiner.py

@@ -0,0 +1,176 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+"""Provides utility to combine a vision backbone with a language backbone."""
+
+from copy import copy
+from typing import List, Optional
+
+import torch
+import torch.nn as nn
+
+from torch.nn.attention import sdpa_kernel, SDPBackend
+
+from .act_ckpt_utils import activation_ckpt_wrapper
+from .necks import Sam3DualViTDetNeck
+
+
+class SAM3VLBackbone(nn.Module):
+    """This backbone combines a vision backbone and a language backbone without fusion.
+    As such it is more of a convenience wrapper to handle the two backbones together.
+
+    It adds support for activation checkpointing and compilation.
+    """
+
+    def __init__(
+        self,
+        visual: Sam3DualViTDetNeck,
+        text,
+        compile_visual: bool = False,
+        act_ckpt_whole_vision_backbone: bool = False,
+        act_ckpt_whole_language_backbone: bool = False,
+        scalp=0,
+    ):
+        """Initialize the backbone combiner.
+
+        :param visual: The vision backbone to use
+        :param text: The text encoder to use
+        """
+        super().__init__()
+        self.vision_backbone: Sam3DualViTDetNeck = (
+            torch.compile(visual) if compile_visual else visual
+        )
+        self.language_backbone = text
+        self.scalp = scalp
+        # allow running activation checkpointing on the entire vision and language backbones
+        self.act_ckpt_whole_vision_backbone = act_ckpt_whole_vision_backbone
+        self.act_ckpt_whole_language_backbone = act_ckpt_whole_language_backbone
+
+    def forward(
+        self,
+        samples: torch.Tensor,
+        captions: List[str],
+        input_boxes: Optional[torch.Tensor] = None,
+        additional_text: Optional[List[str]] = None,
+    ):
+        """Forward pass of the backbone combiner.
+
+        :param samples: The input images
+        :param captions: The input captions
+        :param input_boxes: If the text contains place-holders for boxes, this
+            parameter contains the tensor containing their spatial features
+        :param additional_text: This can be used to encode some additional text
+            (different from the captions) in the same forward of the backbone
+        :return: Output dictionary with the following keys:
+            - vision_features: The output of the vision backbone
+            - language_features: The output of the language backbone
+            - language_mask: The attention mask of the language backbone
+            - vision_pos_enc: The positional encoding of the vision backbone
+            - (optional) additional_text_features: The output of the language
+                backbone for the additional text
+            - (optional) additional_text_mask: The attention mask of the
+                language backbone for the additional text
+        """
+        output = self.forward_image(samples)
+        device = output["vision_features"].device
+        output.update(self.forward_text(captions, input_boxes, additional_text, device))
+        return output
+
+    def forward_image(self, samples: torch.Tensor):
+        return activation_ckpt_wrapper(self._forward_image_no_act_ckpt)(
+            samples=samples,
+            act_ckpt_enable=self.act_ckpt_whole_vision_backbone and self.training,
+        )
+
+    def _forward_image_no_act_ckpt(self, samples):
+        # Forward through backbone
+        sam3_features, sam3_pos, sam2_features, sam2_pos = self.vision_backbone.forward(
+            samples
+        )
+        if self.scalp > 0:
+            # Discard the lowest resolution features
+            sam3_features, sam3_pos = (
+                sam3_features[: -self.scalp],
+                sam3_pos[: -self.scalp],
+            )
+            if sam2_features is not None and sam2_pos is not None:
+                sam2_features, sam2_pos = (
+                    sam2_features[: -self.scalp],
+                    sam2_pos[: -self.scalp],
+                )
+
+        sam2_output = None
+
+        if sam2_features is not None and sam2_pos is not None:
+            sam2_src = sam2_features[-1]
+            sam2_output = {
+                "vision_features": sam2_src,
+                "vision_pos_enc": sam2_pos,
+                "backbone_fpn": sam2_features,
+            }
+
+        sam3_src = sam3_features[-1]
+        output = {
+            "vision_features": sam3_src,
+            "vision_pos_enc": sam3_pos,
+            "backbone_fpn": sam3_features,
+            "sam2_backbone_out": sam2_output,
+        }
+
+        return output
+
+    def forward_text(
+        self, captions, input_boxes=None, additional_text=None, device="cuda"
+    ):
+        return activation_ckpt_wrapper(self._forward_text_no_ack_ckpt)(
+            captions=captions,
+            input_boxes=input_boxes,
+            additional_text=additional_text,
+            device=device,
+            act_ckpt_enable=self.act_ckpt_whole_language_backbone and self.training,
+        )
+
+    def _forward_text_no_ack_ckpt(
+        self,
+        captions,
+        input_boxes=None,
+        additional_text=None,
+        device="cuda",
+    ):
+        output = {}
+
+        # Forward through text_encoder
+        text_to_encode = copy(captions)
+        if additional_text is not None:
+            # if there are additional_text, we piggy-back them into this forward.
+            # They'll be used later for output alignment
+            text_to_encode += additional_text
+
+        sdpa_context = sdpa_kernel(
+            [
+                SDPBackend.MATH,
+                SDPBackend.EFFICIENT_ATTENTION,
+                SDPBackend.FLASH_ATTENTION,
+            ]
+        )
+
+        with sdpa_context:
+            text_attention_mask, text_memory, text_embeds = self.language_backbone(
+                text_to_encode, input_boxes, device=device
+            )
+
+        if additional_text is not None:
+            output["additional_text_features"] = text_memory[:, -len(additional_text) :]
+            output["additional_text_mask"] = text_attention_mask[
+                -len(additional_text) :
+            ]
+
+        text_memory = text_memory[:, : len(captions)]
+        text_attention_mask = text_attention_mask[: len(captions)]
+        text_embeds = text_embeds[:, : len(captions)]
+        output["language_features"] = text_memory
+        output["language_mask"] = text_attention_mask
+        output["language_embeds"] = (
+            text_embeds  # Text embeddings before forward to the encoder
+        )
+
+        return output