added freeze weights for GPT-2

run_glue.py

@@ -202,6 +202,24 @@ class ModelArguments:
         metadata={"help": "Will enable to load a pretrained model whose head dimensions are different."},
     )
     
+    freeze_weights: bool = field(
+        default=False,
+        metadata={"help": "Freeze encoder weights"},
+    )
+
+
+def freeze_model_weights(model: torch.nn.Module) -> None:
+    count = 0
+    print(len(model.parameters())
+    for param in model.parameters():
+        count += 1
+        if count <= 10:
+            logger.info(f'Freezing layer {count}')
+            param.requires_grad = False
+        else:
+            logger.info(f'Ignoring layer {count}')
+
+
 
 def main():
     # See all possible arguments in src/transformers/training_args.py
@@ -376,6 +394,11 @@ def main():
         ignore_mismatched_sizes=model_args.ignore_mismatched_sizes,
     )
     
+    if model_args.freeze_weights:
+        logger.info("Freezing encoder weights")
+        freeze_model_weights(model.decoder)
+        
+
     # Preprocessing the raw_datasets
     if data_args.task_name is not None:
         sentence1_key, sentence2_key = task_to_keys[data_args.task_name]