Prepare optimizer only when args.do_train is True

examples/lm_finetuning/simple_lm_finetuning.py

@@ -534,6 +534,7 @@ def main():
         model = torch.nn.DataParallel(model)
 
     # Prepare optimizer
+    if args.do_train:
         param_optimizer = list(model.named_parameters())
         no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
         optimizer_grouped_parameters = [

examples/run_classifier.py

@@ -763,6 +763,7 @@ def main():
         model = torch.nn.DataParallel(model)
 
     # Prepare optimizer
+    if args.do_train:
         param_optimizer = list(model.named_parameters())
         no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
         optimizer_grouped_parameters = [

examples/run_openai_gpt.py

@@ -183,6 +183,7 @@ def main():
     eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)
 
     # Prepare optimizer
+    if args.do_train:
         param_optimizer = list(model.named_parameters())
         no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
         optimizer_grouped_parameters = [

examples/run_squad.py

@@ -922,6 +922,7 @@ def main():
         model = torch.nn.DataParallel(model)
 
     # Prepare optimizer
+    if args.do_train:
         param_optimizer = list(model.named_parameters())
 
         # hack to remove pooler, which is not used

examples/run_swag.py

@@ -385,6 +385,7 @@ def main():
         model = torch.nn.DataParallel(model)
 
     # Prepare optimizer
+    if args.do_train:
         param_optimizer = list(model.named_parameters())
 
         # hack to remove pooler, which is not used