HOW CAN I Download vit_huge_patch14_224_in21k.pth

[twisti14]

I try to use your VIT model to train my dataset,my value acc is about 82%,but I want to use vit_huge_patch14_224_in21k.pth or Modify the patch size directly，I don't know if it's feasible.By the way,my dataset is about bridge damage.

[dev-jinwoohong]

You can download and use the model with the code provided below.

import torch
import timm

model = timm.create_model('vit_huge_patch14_224.orig_in21k', pretrained=True)
model.eval()

input_tensor = torch.randn(1, 3, 224, 224)

with torch.no_grad():
    output = model(input_tensor)

print(output.shape)

[rwightman]

@twisti14 As per @dev-jinwoohong response, that model was remapped and it's loaded in current timm releases as per that snippet. If you need to download manually you can find the url / hf hub entry 'timm/vit_huge_patch14_224.orig_in21k' and download the .bin or .safetensors file manually

All that said, this specific weight is not that great compared to others. If your dataset is 'bridge damage' it's likely not very big, so you'd get much better results using a smaller model trained on more data, like fine-tuning one of the contrastive pretrained laion / dfn / siglip models in the base, large, so400m size range.

[rwightman]

Maybe even my sbb in12k pretrained weights would be easier / better to fine-tune on this than that old huge model. https://huggingface.co/collections/timm/searching-for-better-vit-baselines-663eb74f64f847d2f35a9c19

[twisti14]

Thanks for your reply. I will have a try.