I would expect to exclude the class_token entirely from masking and from mask calculation. Therefore, the output should have shapes:
idx_keep.shape: (batch_size, 65)
idx_mask.shape: (batch_size, 192)
Additionally, the return type annotation seems to be outdated (torch.Tensor instead of Tuple[torch.Tensor, torch.Tensor])
Thank you in advance, and especially thank you for this fantastic repo.
random_token_mask from models.utils masks a patch too much when mask_class_token=False.
Input: size = (batch_size, sequence_length) <- sequence length includes class token mask_ratio = 0.75 mask_class_token = False
Output shapes: idx_keep.shape: (batch_size, 64) idx_mask.shape: (batch_size, 193)
I would expect to exclude the class_token entirely from masking and from mask calculation. Therefore, the output should have shapes: idx_keep.shape: (batch_size, 65) idx_mask.shape: (batch_size, 192)
Additionally, the return type annotation seems to be outdated (torch.Tensor instead of Tuple[torch.Tensor, torch.Tensor])
Thank you in advance, and especially thank you for this fantastic repo.