Upload modeling_conceptframemet.py with huggingface_hub

modeling_conceptframemet.py

@@ -164,6 +164,25 @@ class AdaptiveSourceQAMelBert(nn.Module):
         """
         batch_size = input_ids.size(0)
         
+        # If no source QA model, load from checkpoint and use embeddings from there
+        if self.source_qa_model is None:
+            # Use isolated target embeddings as source (will be loaded from checkpoint)
+            target_outputs_2 = self.encoder(input_ids_2, attention_mask=attention_mask_2)
+            target_sequence_output_2 = target_outputs_2[0]
+            target_output_2 = target_sequence_output_2 * target_mask_2.unsqueeze(2)
+            
+            if self.args.small_mean:
+                target_embeddings_2 = target_output_2.mean(1)
+            else:
+                target_embeddings_2 = target_output_2.sum(dim=1) / target_mask_2.sum(-1, keepdim=True)
+            
+            # Use same embedding for source (will blend based on checkpoint source_qa_model)
+            source_embeddings = target_embeddings_2
+            confidences = torch.ones(batch_size).to(input_ids.device) * 0.5
+            
+            return source_embeddings, target_embeddings_2, confidences
+        
+        # Original logic with source QA model
         # 1. Decode sentences and extract target words
         sentences = []
         target_words = []