Figure 4 in paper may be wrong according to the code?

[XuanxuanGao]

According to the code, the Figure 4. Left Stage 1 in paper should drop the Linear Embedding and take the Patch Embedding inside?

Looking forward to your reply~

[CarpeDiemly]

我也发现了这个问题，代码里用了if else语句，所以在第一个阶段，patch embedding和linear embedding只能存在一个，而不像Fig.4里画的那样串行。

[mywebinfo65536]

Yes, I have the same question.

[drkostas]

I have the same question. It's even more confusing because in this code snippet:

if i == 0:
     patch_embed = Head(num_conv)
else:
     patch_embed = OverlapPatchEmbed(
                                    img_size=img_size if i == 0 else img_size // (2 ** (i + 1)),
                                    patch_size=7 if i == 0 else 3,
                                    stride=4 if i == 0 else 2,
                                    in_chans=in_chans if i == 0 else embed_dims[i - 1],
                                    embed_dim=embed_dims[i])

the individual if i == 0 else parts of the OverlapPatchEmbed definition are never used since this part is only called when i != 0.

[OliverRensu]

In the first stage, we use conv stem (Head) to replace naive non-overlap patch embedding in ViT. In the following stage, we use overlap patchembed.

[OliverRensu]

You can regard the Head is the combination of patch embedding and Linear Embedding

[drkostas]

Thanks for your reply. Since I can regard it as a combination of patch and linear embedding, is there a way to calculate the number of patches like how it's calculated in the OverlapPatchEmbed?

[mywebinfo65536]

yes, I also have the question that how can we get the patch number in your Head (of the Stage1) ?