minor update use_permute_fix

Author: loribonna

models/dualprompt.py

@@ -27,7 +27,7 @@ def get_parser(parser) -> ArgumentParser:
         parser.set_defaults(optimizer='adam', batch_size=128)
 
         parser.add_argument('--pretrained', default=1, type=binary_to_boolean_type, help='Load pretrained model or not')
-        parser.add_argument('--use_fix_permute', type=binary_to_boolean_type, default=0, help='Apply fix to reshape issue from original implementation (ref: issue #56)')
+        parser.add_argument('--use_permute_fix', type=binary_to_boolean_type, default=0, help='Apply fix to reshape issue from original implementation (ref: issue #56)')
 
         # Optimizer parameters
         parser.add_argument('--clip_grad', type=float, default=1.0, metavar='NORM', help='Clip gradient norm (default: None, no clipping)')

models/dualprompt_utils/model.py

@@ -14,6 +14,7 @@ def __init__(self, args: Namespace, n_classes: int):
             num_classes=n_classes,
             drop_rate=0,
             drop_path_rate=0,
+            args=args
         )
         self.original_model.eval()
 
@@ -41,6 +42,7 @@ def __init__(self, args: Namespace, n_classes: int):
             e_prompt_layer_idx=args.e_prompt_layer_idx,
             use_prefix_tune_for_e_prompt=args.use_prefix_tune_for_e_prompt,
             same_key_value=args.same_key_value,
+            args=args
         )
 
         if args.freeze:

models/dualprompt_utils/prompt.py

@@ -120,10 +120,10 @@ def forward(self, x_embed, prompt_mask=None, cls_features=None):
                 num_layers, dual, batch_size, top_k, length, num_heads, heads_embed_dim = batched_prompt_raw.shape
                 if self.use_permute_fix:  # this fixes issue #56
                     batched_prompt_raw = batched_prompt_raw.permute(0, 2, 1, 3, 4, 5, 6)
-                else:  # this follows the original implementation from https://github.dev/google-research/l2p
-                    batched_prompt = batched_prompt_raw.reshape(
-                        num_layers, batch_size, dual, top_k * length, num_heads, heads_embed_dim
-                    )
+                # else this follows the original implementation from https://github.dev/google-research/l2p
+                batched_prompt = batched_prompt_raw.reshape(
+                    num_layers, batch_size, dual, top_k * length, num_heads, heads_embed_dim
+                )
             else:
                 batched_prompt_raw = self.prompt[:, idx]
                 num_layers, batch_size, top_k, length, embed_dim = batched_prompt_raw.shape