修改了验证集正确率问题，之前打印验证集正确率100%，实际测试并不是100%

faiss_vector_db/build_faiss_index.py

@@ -499,7 +499,7 @@ def main():
     # MODEL_PATH = "../model/ProcessedIngredientRecognition/embedding_20251103_172012/best_embedding_model.pth"
     # MODEL_PATH = "../model/WholeIngredientRecognition/cosface_20251106_134718/best_cosface_model.pth"
     # MODEL_PATH = "../model/DishClassification/cosface_20251105_200551/best_embedding_model.pth"
-    MODEL_PATH = "../model/DishClassification/cosface_20251110_144822/best_cosface_model.pth"
+    MODEL_PATH = "../model/DishClassification/cosface_20251111_094827/best_cosface_model.pth"
     # TRAIN_DIR = "../dataset/ProcessedIngredientRecognition/train"
     # TRAIN_DIR = "../dataset/WholeIngredientRecognition/train"
     TRAIN_DIR = "../dataset/DishClassification/train"

train/train_cosface_embedding.py

@@ -187,9 +187,12 @@ def run_epoch(model, head, loader, criterion, optimizer=None, train: bool = True
         model.eval(); head.eval()
 
     running_loss = 0.0
-    running_acc = 0.0
     n_batches = 0
     
+    # 改为统计总样本数和正确样本数（全局准确率）
+    total_samples = 0
+    correct_samples = 0
+
     with torch.set_grad_enabled(train):
         pbar = tqdm(loader, desc='训练中' if train else '验证中')
         for images, labels in pbar:
@@ -202,17 +205,28 @@ def run_epoch(model, head, loader, criterion, optimizer=None, train: bool = True
             else:
                 logits = head(feats)  # 不加边距
                 loss = criterion(logits, labels)
-            acc = accuracy_top1(logits, labels)
             
+            # 统计样本级别的正确数
+            pred = logits.argmax(dim=1)
+            correct = (pred == labels).sum().item()
+            batch_size = labels.size(0)
+            
+            total_samples += batch_size
+            correct_samples += correct
             running_loss += loss.item()
-            running_acc += acc
             n_batches += 1
+            
+            # 计算当前的全局准确率
+            current_acc = 100.0 * correct_samples / total_samples
+            
             pbar.set_postfix({
                 'Loss': f'{running_loss / n_batches:.4f}',
-                'Acc': f'{running_acc / n_batches:.2f}%'
+                'Acc': f'{current_acc:.2f}%'
             })
 
-    return running_loss / max(n_batches, 1), running_acc / max(n_batches, 1)
+    avg_loss = running_loss / max(n_batches, 1)
+    global_acc = 100.0 * correct_samples / max(total_samples, 1)
+    return avg_loss, global_acc
 
 
 def collect_max_cos_scores(model, head, loader) -> torch.Tensor: