Upload modeling_conceptframemet.py with huggingface_hub

modeling_conceptframemet.py

@@ -288,33 +288,39 @@ class ConceptFrameMetForMetaphorDetection(nn.Module):
         }
     
     @classmethod
-    def from_pretrained(cls, model_path, **kwargs):
+    def from_pretrained(cls, model_path, load_source_from_checkpoint=True, **kwargs):
         """Load model from pretrained checkpoint"""
         # Load config
         config_path = os.path.join(model_path, "config.json")
         with open(config_path, 'r') as f:
             config = json.load(f)
         
-        # Initialize model
-        model = cls(**kwargs)
-        
-        # Load weights
+        # Load weights first to check what's in checkpoint
         weights_path = os.path.join(model_path, "pytorch_model.bin")
-        if os.path.exists(weights_path):
-            state_dict = torch.load(weights_path, map_location='cpu')
-            
-            # Filter out source_qa_model and frame_qa_model weights if model doesn't have them
-            filtered_state_dict = {}
-            for key, value in state_dict.items():
-                # Skip source_qa_model weights if we don't have it
-                if key.startswith('source_qa_model.') and not model.has_source_predictor:
-                    continue
-                # Skip frame_qa_model weights if we don't have it  
-                if key.startswith('frame_qa_model.') and not model.has_frame_predictor:
-                    continue
-                filtered_state_dict[key] = value
-            
-            model.load_state_dict(filtered_state_dict, strict=False)
+        state_dict = torch.load(weights_path, map_location='cpu')
+        
+        # Check if checkpoint has source_qa_model
+        has_source_in_checkpoint = any(k.startswith('source_qa_model.') for k in state_dict.keys())
+        
+        # Initialize model WITHOUT loading external models if they're in checkpoint
+        if has_source_in_checkpoint and load_source_from_checkpoint:
+            # Don't load source from HF, we'll load from checkpoint
+            model = cls(source_qa_model_name=None, **kwargs)
+            # Manually set has_source_predictor since weights are in checkpoint
+            model.has_source_predictor = True
+        else:
+            model = cls(**kwargs)
+        
+        # Load ALL weights from checkpoint (including source_qa_model if present)
+        # Only filter out frame_qa_model if we don't have it
+        filtered_state_dict = {}
+        for key, value in state_dict.items():
+            # Only skip frame_qa_model weights if we don't have it  
+            if key.startswith('frame_qa_model.') and not model.has_frame_predictor:
+                continue
+            filtered_state_dict[key] = value
+        
+        model.load_state_dict(filtered_state_dict, strict=False)
         
         return model