timm.create_model loads the model from a specified local path

[laogonggong847]

hello @rwightman，
Thank you for all the work you've done, and now I have a question.

        self.feature_extractor = timm.create_model(
            backbone,
            pretrained=True,
            features_only=True,
            exportable=True,
            out_indices=self.idx,
        )

Where backbone: str = "wide_resnet_50_2", This code will tell the program to look for the model in question under my C:\Users\user\.cache\torch\hub\checkpoints (which is the path under windows) and if it doesn't exist then download it from the web.

I set the following two parameters in create_model
pretrained=True, features_only=True,

Since "features_only=True", I can't change the program to pretrained=Fasle, and I can't add checkpoint_path = "my_model_path" for the program to go to the folder I specified to find the corresponding model.

So I chose to change OS["TORCH_HOME"] = "New_path"

But I found that when I made this change, the path where timm ends up looking for the model becomes "New_path/hub/checkpoints".

How should I get timm to go to the folder path I specified to find the corresponding model with pretrained=True, features_only=True. And make sure that the timm is looking exactly at the path I provide it with, and not adding subfolders to the path I provide as it would if I were modifying OS["TORCH_HOME"]!

Very much looking forward to your reply, thanks!

[rwightman]

@laogonggong847 functionality was added for this scenario, see my ansewr here #1826 (comment) ... it should work fine with features_only=True

[laogonggong847]

@laogonggong847 functionality was added for this scenario, see my ansewr here #1826 (comment) ... it should work fine with features_only=True

Hello @rwightman
Thank you for your reply, but I noticed that you are not sure what to say in #
in which you state that the version of timm should be greater than 0.9.x, but due to the version requirements of some of the dependent libraries my version of timm is now 0.6.12. Is there a better way to solve this problem other than upgrading the timm?

[rwightman]

@laogonggong847 that functionality was added in 0.9 to solve this use case that a few people had. There is no way of doing it in 0.6.x other than trying to figure out exactly where the torch cache location for the checkpoint is and putting yours in that location. You can't easily avoid the subfolders or redirect without hacking the code...

[rwightman]

You should try 0.9 though as it's quite likely to work even if there is a 0.6.x dependency, effort was made to allow backwards compatibility for majority of changes and the rest should be minor import module name changes (mentioned at top of README).

[rwightman]

Moving to discussion for future ref

[DDUFlyme]

you can use pretrained_cfg as follows:

semantic_model = timm.create_model(
                resnet50, 
                pretrained=True, 
                features_only=True, 
                pretrained_cfg={'file': 'path_to_your_local_model'})

[tdmytrykiv]

Better to use pretrained_cfg_overlay to not overwrite other config keys.
Example:

 model = timm.create_model(
                model_name,
                ...,
                pretrained_cfg_overlay={'file': 'path_to_your_local_model'}
            )

[rwightman]

Better to use pretrained_cfg_overlay to not overwrite other config keys. Example:

 model = timm.create_model(
                model_name,
                ...,
                pretrained_cfg_overlay={'file': 'path_to_your_local_model'}
            )

This is the way for most users. Setting pretrained_cfg_overlay will merge just the contents of the overlay dict with the original pretrasined cfg for the selected model, which is usually what you want. If you set pretrained_cfg it overwrites the full config so you need to ensure it has all the values you might want, mean/std, input_size, etc ...

[rwightman]

If it's not obvious, whenever the 'file' key is set in a pretrained cfg that will override the url or hf_hub_id fields for remote checkpoint locations. You can also use the overlay to point to different remote locations as well though, which can be handy if you see a model you want on the huggingface hub that is in timm format (but not a default model in the library) and doesn't have a proper config (ie just a .bin/.pth/.safetensors file)