Why MobileViT neither loses the patch order nor thespatial order of pixels within each patch?

[ShoufaChen]

Hi, Thank you for the great work.

I am sorry I don't understand why does MobileViT neither loses the patch order nor the spatial order of pixels within each patch?

In Figure 4, I think blue blocks are still permutation equivariant.

Would you mind giving an explanation?

[WYHZQ]

By reading the author's paper, I have the same question.

[sacmehta]

MobileViT unfolds an input with shape [Batch, Channels, Height, Width] into [Batch, Number of pixels per patch, number of patches, Channels] and learn global representations using transformers. The output of transformer has shape of [Batch, Number of pixels per patch, number of patches, Channels]. The folding operation applied on this resulting tensor produces an output of shape [Batch, Channels, Height, Width].

Unlike transformers, MobileViT does not compress the number of pixels per patch. Therefore, after folding operation, we are able to produce the output of the same size as the input.