SpViT on Swin Transformer

[fbragman]

Hi,

Thanks a lot for your contribution!

I was wondering if you could release the SpViT implementation for the Swin Transformer?

Many thanks!

[PeiyanFlying]

Yes, theoretically, this working flow can also be used for Swin Transformer. At this time, we do not have enough time working on the related implementation. We would release in the future If we can make time.

Thanks for support and understanding.

[fbragman]

Thank you for the quick response!

Do you plan on releasing the code used to generate results in Table 3 such as SPViT vs Swin-S and SPViT vs Swin-T?

I am currently implementing your method for my own Swin code. I would be grateful if you could help on the following comments.

In terms of the implementation, the ViT code for instance has

policy = torch.ones(B, init_n + 1, 1, dtype=x.dtype, device=x.device)

When you used Swin - did you generate a new policy for each layer since the number of patches changes for each successive layer?

Also - it is stated that the Token Selector is applied after PatchMerging for layers 2, 3, 4 of Swin. How does this affect how patch merging is performed if token packaging is applied?

Lastly - just to confirm my understanding of the code - this code snippet below corresponds to the token packaging?

              x2 = spatial_x * hard_drop_decision  # placehoder score [96, 196, 384]
              x2_sum = torch.sum(x2, dim=1)  # sum by the N dimension, output (B,N,C)-->(B,C) [96, 384]
              drop_len = torch.sum(hard_drop_decision)
              represent_token = (x2_sum/drop_len).reshape(B, 1, -1)
              x = torch.cat((x,represent_token), dim=1)

whilst the policy tensor is used in the attention module to ensure that the softmax is only done on tokens that are kept / repackaged?

Thanks for your guidance!