Fix hang: cache images in memory before delete, use 2 workers, vis from cache

app.py

@@ -46,7 +46,7 @@ dinov2_extractor = None
 ocr_engine = None
 bge_tokenizer = None
 bge_model = None
-executor = ThreadPoolExecutor(max_workers=1)
+executor = ThreadPoolExecutor(max_workers=2)
 
 session_data = {
     'startup_stats': None,
@@ -309,35 +309,45 @@ def _run_detect_pipeline(uploaded_paths, uname, ocr_engine, bge_tokenizer,
         if log['is_same_scene'] and log['query_path'] in uploaded_paths:
             same_scene_paths.add(log['query_path'])
 
-    # ★★★ 关键修复：先构建可视化，再删除文件 ★★★
+    # ★ 先把需要可视化的图片数据读入内存，再删除文件 ★
+    # 这样可视化构建不依赖磁盘文件，且不会被删除操作影响
+    _image_cache = {}
+    for log in evaluation_logs:
+        if not log.get('is_same_scene'):
+            continue
+        for key in ('query_path', 'history_path'):
+            p = log.get(key, '')
+            if p and p not in _image_cache and os.path.exists(p):
+                try:
+                    import cv2, numpy as np
+                    data = np.fromfile(p, dtype=np.uint8)
+                    img = cv2.imdecode(data, cv2.IMREAD_COLOR)
+                    if img is not None:
+                        _image_cache[p] = img
+                except Exception:
+                    pass
+
+    # 删除同一场景的文件
+    for path in same_scene_paths:
+        delete_file(path)
+
+    # 构建可视化（使用内存中的图片数据）
     visualization_data = []
     for log in evaluation_logs:
-        if log.get('is_same_scene'):
-            scene_type = log.get('scene_type', 'complex')
+        if not log.get('is_same_scene'):
+            continue
+        scene_type = log.get('scene_type', 'complex')
+        try:
             if scene_type == 'complex':
-                patch_info = log.get('patch_match_info')
-                if patch_info is None and dinov2_extractor is not None:
-                    # patch_match_info 可能为 None（dinov2_sim_override 模式下）
-                    # 重新用模型推理获取 patch 匹配
-                    q_path = log.get('query_path', '')
-                    h_path = log.get('history_path', '')
-                    if os.path.exists(q_path) and os.path.exists(h_path):
-                        try:
-                            patch_info = dinov2_extractor.compute_patch_matches(q_path, h_path, top_k=50)
-                        except Exception as e:
-                            print(f"  [可视化] 重新计算 patch 匹配失败: {e}")
-                if patch_info:
-                    vis = _build_complex_visualization(log, patch_info)
-                    if vis:
-                        visualization_data.append(vis)
+                vis = _build_complex_vis_from_cache(log, _image_cache)
+                if vis:
+                    visualization_data.append(vis)
             elif scene_type == 'text':
-                vis = _build_text_visualization(log)
+                vis = _build_text_vis_from_cache(log, _image_cache)
                 if vis:
                     visualization_data.append(vis)
-
-    # 现在才删除同一场景的文件
-    for path in same_scene_paths:
-        delete_file(path)
+        except Exception as e:
+            print(f"  [可视化] 生成失败: {e}")
 
     per_image_results = {}
     for orig, path in [(os.path.basename(p), p) for p in uploaded_paths]:
@@ -487,16 +497,10 @@ async def detect(
     }
 
 
-def _build_complex_visualization(log_entry, patch_info):
-    """复杂场景可视化：DINOv2 patch 匹配连线和矩形框
+def _build_complex_vis_from_cache(log_entry, image_cache):
+    """复杂场景可视化：DINOv2 patch 匹配连线
 
-    坐标映射原理：
-    DINOv2 预处理流程：原始图 → Resize(518,短边) → CenterCrop(518)
-    Patch 是在 518x518 的图像上以 14px 为步长划分的网格。
-    要将 patch 坐标映射回原始图像：
-    1. patch_idx → (row, col) → 像素坐标 (518x518 空间)
-    2. 加上 CenterCrop 偏移 → Resize 后图像坐标
-    3. 除以 Resize 缩放因子 → 原始图像坐标
+    图片数据从 image_cache 中获取（内存），不依赖磁盘文件。
     """
     try:
         import cv2
@@ -504,20 +508,18 @@ def _build_complex_visualization(log_entry, patch_info):
 
         q_path = log_entry.get('query_path', '')
         h_path = log_entry.get('history_path', '')
+        patch_info = log_entry.get('patch_match_info')
 
-        if not os.path.exists(q_path) or not os.path.exists(h_path):
-            return None
-
-        img1 = cv2.imdecode(np.fromfile(q_path, dtype=np.uint8), cv2.IMREAD_COLOR)
-        img2 = cv2.imdecode(np.fromfile(h_path, dtype=np.uint8), cv2.IMREAD_COLOR)
+        img1 = image_cache.get(q_path)
+        img2 = image_cache.get(h_path)
         if img1 is None or img2 is None:
             return None
+        if patch_info is None:
+            return None
 
-        # 原始图片尺寸
         orig_h1, orig_w1 = img1.shape[:2]
         orig_h2, orig_w2 = img2.shape[:2]
 
-        # 缩放显示到统一高度
         max_h = 400
         scale1 = max_h / orig_h1
         scale2 = max_h / orig_h2
@@ -533,7 +535,6 @@ def _build_complex_visualization(log_entry, patch_info):
         q_orig_size = patch_info.get('query_image_size', (orig_w1, orig_h1))
         h_orig_size = patch_info.get('hist_image_size', (orig_w2, orig_h2))
 
-        # CenterCrop 偏移和 Resize 缩放（query 和 hist 各自独立）
         q_crop = patch_info.get('query_crop_offset', (0, 0))
         h_crop = patch_info.get('hist_crop_offset', (0, 0))
         q_resize = patch_info.get('query_resize_size', (518, 518))
@@ -541,11 +542,9 @@ def _build_complex_visualization(log_entry, patch_info):
 
         q_n_h, q_n_w = q_grid
         h_n_h, h_n_w = h_grid
-        input_size = 518
         patch_size = 14
 
-        # 计算 Resize 缩放因子：原始图 → Resize 后图的缩放
-        q_scale_resize = q_resize[1] / q_orig_size[1]  # resize_h / orig_h
+        q_scale_resize = q_resize[1] / q_orig_size[1]
         h_scale_resize = h_resize[1] / h_orig_size[1]
 
         gap = 20
@@ -553,41 +552,20 @@ def _build_complex_visualization(log_entry, patch_info):
         canvas[:h1, :w1] = img1_disp
         canvas[:h2, w1+gap:] = img2_disp
 
-        def _patch_to_center(patch_idx, n_w, crop_offset, scale_resize, display_scale, is_right, gap_offset):
-            """将 patch 索引转换为画布上的中心点像素坐标"""
-            row = patch_idx // n_w
-            col = patch_idx % n_w
-
-            # 在 518x518 输入图像上的像素坐标（patch 中心）
-            cx_518 = (col + 0.5) * patch_size
-            cy_518 = (row + 0.5) * patch_size
-
-            # 加上 CenterCrop 偏移，得到 Resize 后图像的坐标
-            cx_resize = cx_518 + crop_offset[0]
-            cy_resize = cy_518 + crop_offset[1]
-
-            # 反向缩放，得到原始图像坐标
-            cx_orig = cx_resize / scale_resize
-            cy_orig = cy_resize / scale_resize
-
-            # 乘以显示缩放因子，得到画布坐标
-            cx_canvas = int(cx_orig * display_scale)
-            cy_canvas = int(cy_orig * display_scale)
-
+        def _patch_center(idx, n_w, crop_off, scale_r, disp_s, is_right, gap_off):
+            row, col = idx // n_w, idx % n_w
+            cx = int(((col + 0.5) * patch_size + crop_off[0]) / scale_r * disp_s)
+            cy = int(((row + 0.5) * patch_size + crop_off[1]) / scale_r * disp_s)
             if is_right:
-                cx_canvas += gap_offset
-
-            return cx_canvas, cy_canvas
+                cx += gap_off
+            return cx, cy
 
         for qi, hi, sim in matches:
-            q_cx, q_cy = _patch_to_center(qi, q_n_w, q_crop, q_scale_resize, scale1, False, 0)
-            h_cx, h_cy = _patch_to_center(hi, h_n_w, h_crop, h_scale_resize, scale2, True, w1 + gap)
+            q_cx, q_cy = _patch_center(qi, q_n_w, q_crop, q_scale_resize, scale1, False, 0)
+            h_cx, h_cy = _patch_center(hi, h_n_w, h_crop, h_scale_resize, scale2, True, w1 + gap)
 
-            # 颜色：相似度越高越绿，越低越黄
             intensity = min(1.0, max(0.0, (sim - 0.3) / 0.7))
             color = (0, int(200 * intensity + 55), int(255 * (1 - intensity)))
-
-            # 只绘制连线
             cv2.line(canvas, (q_cx, q_cy), (h_cx, h_cy), color, 1, cv2.LINE_AA)
 
         n_matches = len(matches)
@@ -614,25 +592,22 @@ def _build_complex_visualization(log_entry, patch_info):
         return None
 
 
-def _build_text_visualization(log_entry):
-    """文本场景可视化：两张图并排 + BGE 相似度 + 关键词统计
+def _build_text_vis_from_cache(log_entry, image_cache):
+    """文本场景可视化：两张图并排 + 标注
 
-    上方：两张图并排缩略图，标注 Query / History
-    下方：BGE 语义相似度 + 共同/差异关键词统计（用 PIL 绘制中文）
+    图片数据从 image_cache 中获取（内存），不依赖磁盘文件。
     """
     try:
         import cv2
         import numpy as np
         from PIL import Image, ImageDraw, ImageFont
+        import re
 
         q_path = log_entry.get('query_path', '')
         h_path = log_entry.get('history_path', '')
 
-        if not os.path.exists(q_path) or not os.path.exists(h_path):
-            return None
-
-        img1 = cv2.imdecode(np.fromfile(q_path, dtype=np.uint8), cv2.IMREAD_COLOR)
-        img2 = cv2.imdecode(np.fromfile(h_path, dtype=np.uint8), cv2.IMREAD_COLOR)
+        img1 = image_cache.get(q_path)
+        img2 = image_cache.get(h_path)
         if img1 is None or img2 is None:
             return None
 
@@ -640,15 +615,12 @@ def _build_text_visualization(log_entry):
         text2 = log_entry.get('history_text', '')
         text_sim = log_entry.get('text_similarity', 0)
 
-        # 提取关键词
-        import re
         words1 = set(re.findall(r'[a-zA-Z\u4e00-\u9fff]{2,}', text1))
         words2 = set(re.findall(r'[a-zA-Z\u4e00-\u9fff]{2,}', text2))
         common_words = words1 & words2
         diff_words1 = words1 - words2
         diff_words2 = words2 - words1
 
-        # 缩放图片
         max_h = 300
         scale1 = max_h / img1.shape[0]
         scale2 = max_h / img2.shape[0]
@@ -658,16 +630,12 @@ def _build_text_visualization(log_entry):
         h1, w1 = img1_s.shape[:2]
         h2, w2 = img2_s.shape[:2]
 
-        # 构建画布：只放两张图并排
         gap = 10
         total_w = w1 + gap + w2
         canvas = np.ones((max_h, total_w, 3), dtype=np.uint8) * 240
-
-        # 放图片
         canvas[:max_h, :w1] = img1_s
         canvas[:max_h, w1+gap:w1+gap+w2] = img2_s
 
-        # 用 PIL 绘制图片标注
         canvas_rgb = cv2.cvtColor(canvas, cv2.COLOR_BGR2RGB)
         pil_img = Image.fromarray(canvas_rgb)
         draw = ImageDraw.Draw(pil_img)
@@ -680,15 +648,12 @@ def _build_text_visualization(log_entry):
             except Exception:
                 font_large = ImageFont.load_default()
 
-        # 图片标注
         draw.text((10, 8), "Query (OCR text)", fill=(0, 0, 0), font=font_large)
         draw.text((w1+gap+10, 8), "History (OCR text)", fill=(0, 0, 0), font=font_large)
 
         common_list = sorted(common_words)[:20]
 
-        # 转回 OpenCV 格式
         canvas = cv2.cvtColor(np.array(pil_img), cv2.COLOR_RGB2BGR)
-
         _, buffer = cv2.imencode('.jpg', canvas, [cv2.IMWRITE_JPEG_QUALITY, 85])
         img_b64 = base64.b64encode(buffer).decode('utf-8')