qwen3.5 优化

fastAPI_tarot.py

@@ -22,6 +22,7 @@ import re
 import asyncio
 import shutil
 import subprocess
+import ast
 from datetime import datetime
 from typing import Optional, List, Dict, Any
 from contextlib import asynccontextmanager
@@ -288,6 +289,35 @@ def append_to_history(req_type: str, prompt: str, status: str, result_path: str
         print(f"Failed to write history: {e}")
 
 
+def extract_json_from_response(text: str) -> dict:
+    """
+    Robustly extract JSON from text, handling:
+    1. Markdown code blocks (```json ... ```)
+    2. Single quotes (Python dict style) via ast.literal_eval
+    """
+    try:
+        # 1. Try to find JSON block
+        json_match = re.search(r'```json\s*(.*?)\s*```', text, re.DOTALL)
+        if json_match:
+            clean_text = json_match.group(1).strip()
+        else:
+            # Try to find { ... } block if no markdown
+            match = re.search(r'\{.*\}', text, re.DOTALL)
+            if match:
+                clean_text = match.group(0).strip()
+            else:
+                clean_text = text.strip()
+        
+        # 2. Try standard JSON
+        return json.loads(clean_text)
+    except Exception as e1:
+        # 3. Try ast.literal_eval for single quotes
+        try:
+            return ast.literal_eval(clean_text)
+        except Exception as e2:
+            # 4. Fail
+            raise ValueError(f"Could not parse JSON: {e1} | {e2} | Content: {text[:100]}...")
+
 def translate_to_sam3_prompt(text: str) -> str:
     """
     使用 Qwen 模型将中文提示词翻译为英文
@@ -567,13 +597,13 @@ def recognize_card_with_qwen(image_path: str) -> dict:
         
         if response.status_code == 200:
             content = response.output.choices[0].message.content[0]['text']
-            import json
             try:
-                clean_content = content.replace("```json", "").replace("```", "").strip()
-                result = json.loads(clean_content)
+                result = extract_json_from_response(content)
+                result["model_used"] = QWEN_MODEL
                 return result
-            except:
-                return {"raw_response": content}
+            except Exception as e:
+                print(f"JSON Parse Error in recognize_card: {e}")
+                return {"raw_response": content, "error": str(e), "model_used": QWEN_MODEL}
         else:
             return {"error": f"API Error: {response.code} - {response.message}"}
             
@@ -602,13 +632,13 @@ def recognize_spread_with_qwen(image_path: str) -> dict:
         
         if response.status_code == 200:
             content = response.output.choices[0].message.content[0]['text']
-            import json
             try:
-                clean_content = content.replace("```json", "").replace("```", "").strip()
-                result = json.loads(clean_content)
+                result = extract_json_from_response(content)
+                result["model_used"] = QWEN_MODEL
                 return result
-            except:
-                return {"raw_response": content, "spread_name": "Unknown"}
+            except Exception as e:
+                print(f"JSON Parse Error in recognize_spread: {e}")
+                return {"raw_response": content, "error": str(e), "spread_name": "Unknown", "model_used": QWEN_MODEL}
         else:
             return {"error": f"API Error: {response.code} - {response.message}"}
             
@@ -951,6 +981,10 @@ async def recognize_tarot(
     processor = request.app.state.processor
 
     try:
+        # 在执行 GPU 操作前，切换到线程中运行，避免阻塞主线程（虽然 SAM3 推理在 CPU 上可能已经很快，但为了保险）
+        # 注意：processor 内部调用了 torch，如果是在 GPU 上，最好不要多线程调用同一个 model
+        # 但这里只是推理，且是单次请求。
+        # 如果是 CPU 推理，run_in_executor 有助于防止阻塞 loop
         inference_state = processor.set_image(image)
         output = processor.set_text_prompt(state=inference_state, prompt="tarot card")
         masks, boxes, scores = output["masks"], output["boxes"], output["scores"]
@@ -975,15 +1009,25 @@ async def recognize_tarot(
         main_file_path = None
         main_file_url = None
 
-    # Step 0: 牌阵识别
+    # Step 0: 牌阵识别 (异步启动)
     spread_info = {"spread_name": "Unknown"}
+    spread_task = None
     if main_file_path:
         # 使用原始图的一份拷贝给 Qwen 识别牌阵
         temp_raw_path = os.path.join(output_dir, "raw_for_spread.jpg")
         image.save(temp_raw_path)
-        spread_info = recognize_spread_with_qwen(temp_raw_path)
-    
+        # 将同步调用包装为异步任务
+        loop = asyncio.get_event_loop()
+        spread_task = loop.run_in_executor(None, recognize_spread_with_qwen, temp_raw_path)
+
     if detected_count != expected_count:
+        # 如果数量不对，等待牌阵识别完成（如果已启动）再返回
+        if spread_task:
+            try:
+                spread_info = await spread_task
+            except Exception as e:
+                print(f"Spread recognition failed: {e}")
+                
         duration = time.time() - start_time
         append_to_history("tarot-recognize", f"expected: {expected_count}", "failed", result_path=f"results/{request_id}/{main_filename}" if main_file_url else None, details=f"Detected {detected_count}, expected {expected_count}", duration=duration)
         return JSONResponse(
@@ -1005,21 +1049,47 @@ async def recognize_tarot(
         append_to_history("tarot-recognize", f"expected: {expected_count}", "failed", details=f"Crop Error: {str(e)}", duration=duration)
         raise HTTPException(status_code=500, detail=f"抠图处理错误: {str(e)}")
 
-    # 遍历每张卡片进行识别
+    # 遍历每张卡片进行识别 (并发)
     tarot_cards = []
+    
+    # 1. 准备任务列表
+    loop = asyncio.get_event_loop()
+    card_tasks = []
+    
     for obj in saved_objects:
         fname = obj["filename"]
         file_path = os.path.join(output_dir, fname)
-        
-        # 调用 Qwen-VL 识别 (串行)
-        recognition_res = recognize_card_with_qwen(file_path)
-        
+        # 创建异步任务
+        # 使用 lambda 来延迟调用，确保参数传递正确
+        task = loop.run_in_executor(None, recognize_card_with_qwen, file_path)
+        card_tasks.append(task)
+    
+    # 2. 等待所有卡片识别任务完成
+    # 同时等待牌阵识别任务 (如果还在运行)
+    if card_tasks:
+        all_card_results = await asyncio.gather(*card_tasks)
+    else:
+        all_card_results = []
+    
+    if spread_task:
+        try:
+            # 如果之前没有await spread_task，这里确保它完成
+            # 注意：如果 detected_count != expected_count 分支已经 await 过了，这里不会重复执行
+            # 但那个分支有 return，所以这里肯定是还没 await 的
+            spread_info = await spread_task
+        except Exception as e:
+            print(f"Spread recognition failed: {e}")
+
+    # 3. 组装结果
+    for i, obj in enumerate(saved_objects):
+        fname = obj["filename"]
         file_url = str(request.url_for("static", path=f"results/{request_id}/{fname}"))
+        
         tarot_cards.append({
             "url": file_url,
             "is_rotated": obj["is_rotated_by_algorithm"],
             "orientation_status": "corrected_to_portrait" if obj["is_rotated_by_algorithm"] else "original_portrait",
-            "recognition": recognition_res,
+            "recognition": all_card_results[i],
             "note": obj["note"]
         })
 
@@ -1083,14 +1153,26 @@ async def segment_face(
 
     # 调用独立服务进行处理
     try:
-        result = human_analysis_service.process_face_segmentation_and_analysis(
-            processor=processor,
-            image=image,
-            prompt=final_prompt,
-            output_base_dir=RESULT_IMAGE_DIR,
-            qwen_model=QWEN_MODEL,
-            analysis_prompt=PROMPTS["face_analysis"]
-        )
+        # 使用新增加的异步并发函数
+        if hasattr(human_analysis_service, "process_face_segmentation_and_analysis_async"):
+            result = await human_analysis_service.process_face_segmentation_and_analysis_async(
+                processor=processor,
+                image=image,
+                prompt=final_prompt,
+                output_base_dir=RESULT_IMAGE_DIR,
+                qwen_model=QWEN_MODEL,
+                analysis_prompt=PROMPTS["face_analysis"]
+            )
+        else:
+            # 回退到同步
+            result = human_analysis_service.process_face_segmentation_and_analysis(
+                processor=processor,
+                image=image,
+                prompt=final_prompt,
+                output_base_dir=RESULT_IMAGE_DIR,
+                qwen_model=QWEN_MODEL,
+                analysis_prompt=PROMPTS["face_analysis"]
+            )
     except Exception as e:
         import traceback
         traceback.print_exc()