Upload modeling_conceptframemet.py with huggingface_hub

modeling_conceptframemet.py

@@ -50,10 +50,22 @@ class ConceptFrameMetForMetaphorDetection(nn.Module):
         
         # Load encoder (RoBERTa) with correct type_vocab_size
         from transformers import RobertaConfig
-        encoder_config = RobertaConfig.from_pretrained(encoder_model_name)
-        encoder_config.type_vocab_size = 4  # CRITICAL: Match training config
         
-        self.encoder = RobertaModel.from_pretrained(encoder_model_name, config=encoder_config)
+        # Load base model first
+        self.encoder = RobertaModel.from_pretrained(encoder_model_name)
+        
+        # Resize token_type_embeddings to match training (type_vocab_size=4)
+        # This is needed because the model was trained with 4 token types
+        if self.encoder.embeddings.token_type_embeddings.weight.shape[0] != 4:
+            old_embeddings = self.encoder.embeddings.token_type_embeddings
+            new_embeddings = nn.Embedding(4, old_embeddings.embedding_dim)
+            # Copy the original embedding (for type 0)
+            new_embeddings.weight.data[0] = old_embeddings.weight.data[0]
+            # Initialize the rest
+            new_embeddings.weight.data[1:].normal_(mean=0.0, std=self.encoder.config.initializer_range)
+            self.encoder.embeddings.token_type_embeddings = new_embeddings
+            self.encoder.config.type_vocab_size = 4
+        
         self.tokenizer = RobertaTokenizer.from_pretrained(encoder_model_name)
         self.config = self.encoder.config