Loading timm pre-trained model weight from huggingface models hub

[HoseinAkbarzadeh]

/I wanted to load pretrained model with timm but the models is not available via timm create_model. However the pretrained model seems available on huggingface model hub ("timm/vit_base_patch16_384.augreg_in21k_ft_in1k" for example). In the model description on the huggingface it is instructed to load model for timm with create_model and and add "hf_hub:" prefix before model name,
import timm
model = timm.create_model("hf_hub:timm/vit_base_patch16_384.augreg_in21k_ft_in1k", pretrained=True)
but when i try that it returns an error KeyError that seem to me the model couldn't be found.
KeyError Traceback (most recent call last)
vision_transformer.py in _create_vision_transformer(variant, pretrained, **kwargs)
788 pretrained_cfg=pretrained_cfg,
789 pretrained_filter_fn=checkpoint_filter_fn,
--> 790 pretrained_custom_load='npz' in pretrained_cfg['url'],
791 **kwargs)
792 return model
KeyError: 'url'
In general there are some pretrained models on huggingface but that does not appear when timm.list_models is used, I am asking how can I use those models?

[rwightman]

@HoseinAkbarzadeh you should use a newer timm version (0.8+) for fully HF hub support, it was only partially working in 0.6.

pip install --pre timm