Update mask_adapter/sam_maskadapter.py

mask_adapter/sam_maskadapter.py

@@ -239,7 +239,7 @@ class SAMPointVisualizationDemo(object):
         lvis_classes = [x[x.find(':')+1:] for x in lvis_classes]
                 
         self.class_names = thing_classes + stuff_classes + lvis_classes
-        self.text_embedding = torch.from_numpy(np.load("./text_embedding/lvis_coco_text_embedding.npy"))
+        #self.text_embedding = torch.from_numpy(np.load("./text_embedding/lvis_coco_text_embedding.npy"))
     
         self.class_names = self._load_class_names() 
 
@@ -280,7 +280,7 @@ class SAMPointVisualizationDemo(object):
         
         return clip_vis_dense
     
-    def run_on_image_with_points(self, ori_image, points):
+    def run_on_image_with_points(self, ori_image, points,text_features):
         height, width, _ = ori_image.shape
 
         image = ori_image
@@ -311,13 +311,13 @@ class SAMPointVisualizationDemo(object):
             # text_features = self.clip_model.encode_text(text.cuda())
             # text_features /= text_features.norm(dim=-1, keepdim=True)
             #np.save("/home/yongkangli/Mask-Adapter/text_embedding/lvis_coco_text_embedding.npy", text_features.cpu().numpy())
-            text_features = self.text_embedding.to(self.mask_adapter.device)
-            features = self.extract_features_convnext(image.to(self.mask_adapter.device).float())
+            #text_features = self.text_embedding.to(self.mask_adapter.device)
+            features = self.extract_features_convnext(image.to(text_features).float())
             clip_feature = features['clip_vis_dense']
 
             clip_vis_dense = self.visual_prediction_forward_convnext_2d(clip_feature)
 
-            semantic_activation_maps = self.mask_adapter(clip_vis_dense, pred_masks.tensor.unsqueeze(0).to(self.mask_adapter.device).float())
+            semantic_activation_maps = self.mask_adapter(clip_vis_dense, pred_masks.tensor.unsqueeze(0).to(text_features).float())
             maps_for_pooling = F.interpolate(semantic_activation_maps, size=clip_feature.shape[-2:], mode='bilinear', align_corners=False)
 
             B, C = clip_feature.size(0), clip_feature.size(1)