certify error about dmae

[quanlin-wu]

Sorry to answer your question so late. You said that the bug happend when you run certify_cifar10.py. In fact, 768 is the embedding dimension of ViT Base, I think the problem should lies in 'models_vit.py', likely to be the nn.LayerNorm. You need to check the version of 'timm', our version is 0.5.4.

[yuchongy1]

Hi,

Thanks for your reply.

I believe I have

assert timm.__version__ == "0.5.4"  # version check

in my code, which ensures the correct version?

[quanlin-wu]

Can you print the shape of tensor in 'models_vit.py'?
I want to know whether the bug happen before
'''
for blk in self.blocks:
x = blk(x)
'''
or not.

[yuchongy1]

Sorry, I think I have the version check in the main but not in the models_vit.py,
but I suppose the check in the main.py should already ensure the timm is 0.5.4 for models_vit.py?

I put the print statement in Smooth.py, "_sample_noise" function,
just feeding "noisy" into self.base_classifier.

[quanlin-wu]

The bug happened when you want to evaluate the model on cifar10. Did
you run the pre-training or fine-tuning successfullly? If you directly run 'certfy_cifar10.py', the assert code may not take effect.
The shape of 'nosiy' is correct. It seems the problem lies in the norm layers in the model (the base classifier). Checking the shape of tensors in the vit model may help.

[yuchongy1]

The pre-training and finetuning can run without errors. I will check the shape in vit.