2025-07-31 22:05 🚀

models/clip_ebc/convnext.py

@@ -53,7 +53,8 @@ class ConvNeXt(nn.Module):
         self.model_name, self.weight_name = model_name, weight_name
         self.block_size = block_size
 
-        model = open_clip.create_model_from_pretrained(model_name, weight_name, return_transform=False).visual
+        # model = open_clip.create_model_from_pretrained(model_name, weight_name, return_transform=False).visual
+        model = open_clip.create_model(model_name=model_name, pretrained=False, load_weights=False, return_transform=False).visual
 
         self.adapter = adapter
         if adapter:

models/clip_ebc/mobileclip.py

@@ -41,7 +41,8 @@ class MobileCLIP(nn.Module):
         self.model_name, self.weight_name = model_name, weight_name
         self.block_size = block_size
 
-        model = open_clip.create_model_from_pretrained(model_name, weight_name, return_transform=False).visual
+        # model = open_clip.create_model_from_pretrained(model_name, weight_name, return_transform=False).visual
+        model = open_clip.create_model(model_name=model_name, pretrained=False, load_weights=False, return_transform=False).visual
 
         self.adapter = adapter
         if adapter:

models/clip_ebc/resnet.py

@@ -49,7 +49,8 @@ class ResNet(nn.Module):
         self.model_name, self.weight_name = model_name, weight_name
         self.block_size = block_size
 
-        model = open_clip.create_model_from_pretrained(model_name, weight_name, return_transform=False).visual
+        # model = open_clip.create_model_from_pretrained(model_name, weight_name, return_transform=False).visual
+        model = open_clip.create_model(model_name=model_name, pretrained=False, load_weights=False, return_transform=False).visual
 
         self.adapter = adapter
         if adapter:

models/clip_ebc/vit.py

@@ -95,7 +95,8 @@ class ViT(nn.Module):
         self.vpt_drop = vpt_drop
         self.adapter = adapter
 
-        model = open_clip.create_model_from_pretrained(model_name, weight_name, return_transform=False).visual
+        # model = open_clip.create_model_from_pretrained(model_name, weight_name, return_transform=False).visual
+        model = open_clip.create_model(model_name=model_name, pretrained=False, load_weights=False, return_transform=False).visual
 
         # Always freeze the parameters of the model
         for param in model.parameters():

models/ebc/csrnet.py

@@ -27,7 +27,7 @@ class CSRNet(nn.Module):
         self.model_name = model_name
 
         vgg = VGG(make_vgg_layers(encoder_cfg, in_channels=3, batch_norm="bn" in model_name, dilation=1))
-        vgg.load_state_dict(load_state_dict_from_url(vgg_urls[model_name]), strict=False)
+        # vgg.load_state_dict(load_state_dict_from_url(vgg_urls[model_name]), strict=False)
         self.encoder = vgg.features
         self.encoder_reduction = 8
         self.encoder_channels = 512

models/ebc/hrnet.py

@@ -27,7 +27,8 @@ class HRNet(nn.Module):
         self.model_name = model_name
         self.block_size = block_size if block_size is not None else 32
 
-        model = timm.create_model(model_name, pretrained=True)
+        # model = timm.create_model(model_name, pretrained=True)
+        model = timm.create_model(model_name, pretrained=False)
 
         self.conv1 = model.conv1
         self.bn1 = model.bn1

models/ebc/timm_models.py

@@ -151,7 +151,8 @@ class TIMMModel(nn.Module):
         assert model_name in supported_models, f"Backbone {model_name} not supported. Supported models are {supported_models}"
         assert block_size is None or block_size in [8, 16, 32], f"Block size should be one of [8, 16, 32], but got {block_size}."
         self.model_name = model_name
-        self.encoder = create_model(model_name, pretrained=True, features_only=True, out_indices=[-1])
+        # self.encoder = create_model(model_name, pretrained=True, features_only=True, out_indices=[-1])
+        self.encoder = create_model(model_name, pretrained=False, features_only=True, out_indices=[-1])
         self.encoder_channels = self.encoder.feature_info.channels()[-1]
         self.encoder_reduction = self.encoder.feature_info.reduction()[-1]
         self.block_size = block_size if block_size is not None else self.encoder_reduction

models/ebc/vgg.py

@@ -210,42 +210,42 @@ class VGG(nn.Module):
 
 def vgg11() -> VGG:
     model = VGG(make_vgg_layers(vgg_cfgs["A"]))
-    model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg11"]), strict=False)
+    # model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg11"]), strict=False)
     return model
 
 def vgg11_bn() -> VGG:
     model = VGG(make_vgg_layers(vgg_cfgs["A"], batch_norm=True))
-    model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg11_bn"]), strict=False)
+    # model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg11_bn"]), strict=False)
     return model
 
 def vgg13() -> VGG:
     model = VGG(make_vgg_layers(vgg_cfgs["B"]))
-    model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg13"]), strict=False)
+    # model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg13"]), strict=False)
     return model
 
 def vgg13_bn() -> VGG:
     model = VGG(make_vgg_layers(vgg_cfgs["B"], batch_norm=True))
-    model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg13_bn"]), strict=False)
+    # model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg13_bn"]), strict=False)
     return model
 
 def vgg16() -> VGG:
     model = VGG(make_vgg_layers(vgg_cfgs["D"]))
-    model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg16"]), strict=False)
+    # model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg16"]), strict=False)
     return model
 
 def vgg16_bn() -> VGG:
     model = VGG(make_vgg_layers(vgg_cfgs["D"], batch_norm=True))
-    model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg16_bn"]), strict=False)
+    # model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg16_bn"]), strict=False)
     return model
 
 def vgg19() -> VGG:
     model = VGG(make_vgg_layers(vgg_cfgs["E"]))
-    model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg19"]), strict=False)
+    # model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg19"]), strict=False)
     return model
 
 def vgg19_bn() -> VGG:
     model = VGG(make_vgg_layers(vgg_cfgs["E"], batch_norm=True))
-    model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg19_bn"]), strict=False)
+    # model.load_state_dict(state_dict=load_state_dict_from_url(vgg_urls["vgg19_bn"]), strict=False)
     return model
 
 def _vgg_encoder(model_name: str, block_size: Optional[int] = None, norm: str = "none", act: str = "none") -> VGGEncoder:

models/ebc/vit.py

@@ -86,7 +86,8 @@ class ViT(nn.Module):
         self.num_vpt = num_vpt
         self.vpt_drop = vpt_drop
 
-        model = timm.create_model(model_name, pretrained=True)
+        # model = timm.create_model(model_name, pretrained=True)
+        model = timm.create_model(model_name, pretrained=False)
 
         self.input_size = input_size if input_size is not None else model.patch_embed.img_size
         self.pretrain_size = model.patch_embed.img_size