nmndeep / revisiting-at Goto Github PK

Hello,

First of all - thanks for the great work!
After downloading it from the link, I am trying to load a given checkpoint.
However, the load command fails -
torch.load('convnext_b_cvst_robust.pt', map_location='cpu')
and outputs the following error:
RuntimeError: Expected hasRecord("version") to be true, but got false.
More details - I use the same torch version as required, and I have tried several different checkpoints.
How can I resolve this?

Thanks!

Could authors publish the RobustAcc about L∞=8/255 imagenet models? I tried to test the robustAcc of the models. But they are very low.

https://github.com/nmndeep/revisiting-at/blob/932f73f248447addac17542a574ad7bc784c0cbd/main.py#L447-449

Can you please provide in more detail the training recipe used for adversarial finetuning on downstream datasets(cifar10, cifar1000, flowers)? What optimizer, augmentations are used? Is the adversarial training done similar to how it is done on ImageNet or TRADES framework is used?

Can you please provide weights for other models mentioned in Table 2 such as ViT-S , ConvNeXt-T. ViT-B and ConvNeXt-B(from Table 3)?

Thank you for sharing your code. Could you please recommend a pre-trained model suitable for processing 299*299 images on the ImageNet-1K dataset?

Hi,
thank you for your nice work, it's really enlightening. Especially, using clean pre-trained checkpoints on ImageNet-1K and -21K.

I found that you use:
elif modelname == 'vit_s':
model = create_model('vit_small_patch16_224', pretrained=pretrained)
and
elif modelname == 'vit_s_21k':
model = create_model('deit3_small_patch16_224_in21ft1k', pretrained=pretrained)
to load pre-trained checkpoints.

Do you mean "vit_small_patch16_224" for ImageNet-1K pretraining? But I print the model.default_cfg, it shows the model is timm/vit_small_patch16_224.augreg_in21k_ft_in1k. So it actually loads a checkpoint pre-trained on ImageNet-21K?

Thank you for the great work. I am just wondering if you can explain the reason behind the difference between Table 1 and 2. For example, in Table 1, the ViT-S' performance is (60.3, 30.4), while in Table 2 its performance is (61.5, 31.8) where random init and basic augmentation are adopted. I think the difference is that model in Table 1 are pretrained with standard training for 100 epochs while Table 2 use rand init. But if that's the case, why Table 1 is worse than Table 2? Thank you very much!