modify app

app.py

@@ -100,19 +100,28 @@ def gen_semantic_feature(frame_dir, semantic_dir):
 
     output_path = semantic_dir / "semantic.npy"
     if torch.cuda.is_available():
-        features = torch.cuda.FloatTensor(len(sorted_file_names), 768).fill_(0)
+        features = torch.FloatTensor(len(sorted_file_names), 768).fill_(0)
+        features = features.to(device)
+        
+        for idx, file_name in enumerate(sorted_file_names):
+            fpath = frame_dir / file_name
+            image = preprocess(Image.open(fpath)).unsqueeze(0).to(device)          
+            with torch.no_grad():
+                image_features = model.encode_image(image)
+            features[idx] = image_features[0]
+        features = features.cpu().numpy()
+        np.save(output_path, features)
     else:
         features = torch.FloatTensor(len(sorted_file_names), 768).fill_(0)
-
-    for idx, file_name in enumerate(sorted_file_names):
-        fpath = frame_dir / file_name
-        image = preprocess(Image.open(fpath)).unsqueeze(0).to(device)          
-        with torch.no_grad():
-            image_features = model.encode_image(image)
-        features[idx] = image_features[0]
-    features = features.cpu().numpy()
-    np.save(output_path, features)
-
+        for idx, file_name in enumerate(sorted_file_names):
+            fpath = frame_dir / file_name
+            image = preprocess(Image.open(fpath)).unsqueeze(0).to(device)          
+            with torch.no_grad():
+                image_features = model.encode_image(image)
+            features[idx] = image_features[0]
+        features = features.numpy()
+        np.save(output_path, features)
+        
 def gen_emotion_feature(frame_dir, emotion_dir):
     device = "cuda:0" if torch.cuda.is_available() else "cpu"
     model, preprocess = clip.load("ViT-L/14@336px", device=device)