Closed NightMachinery closed 6 months ago
Yes, we used the ViT implementation in the timm library (ver. 0.5.4), so I think that's possible unless the parameter naming convention of timm library has changed since the version: https://github.com/suinleelab/vit-shapley/blob/master/vit_shapley/modules/vision_transformer.py
@chanwkimlab Thanks! There are two files for each model:
https://aimslab.cs.washington.edu/vitshapley/checkpoints/ImageNette_vit_base_patch16_224_surrogate_3i6zzjnp.ckpt
https://aimslab.cs.washington.edu/vitshapley/checkpoints/ImageNette_vit_base_patch16_224_surrogate_3i6zzjnp_state_dict.ckpt
Am I correct in assuming that I only need to download the state_dict.ckpt
files? Which is loadable using torch.load
?
Can the surrogate weights be loaded in timm's ViT? Which library is the model's parameter naming convention from?