How can i change the path of pre-train models cache?

[ShawYuan97]

I want to change the default cache path

[yohann84L]

To load a pretrained model timm uses the function load_state_dict_from_url() from torch.hub module. In the documentation of this function here, you can see that the default path can be retrieved and set using:

import torch
# Get current default folder
print(torch.hub.get_dir())
>>> ~/.cache/torch/hub

# Set to another default folder
torch.hub.set_dir('your/cache/folder')

You can also set the environment variable TORCH_HOME as it will point to $TORCH_HOME/hub if set.

[ShawYuan97]

thx

[ziimiin14]

This doesn't work anymore for latest timm package. It still always saved in the ~/.cache/huggingface/hub even though i explicitly used the torch.hub.set_dir to other directory

[ziimiin14]

Reference: https://huggingface.co/docs/transformers/v4.38.1/en/installation#cache-setup

For version timm==0.9.16, you need to explicitly set HUGGINGFACE_HUB_CACHE environment variable to the dir you want to store the downloaded pretrained weights.

You can either run the following options:
export HUGGINGFACE_HUB_CACHE=./your/cache/dir

or

import os
os.environ["HUGGINGFACE_HUB_CACHE"] = "./your/cache/dir"