Issue about the padding when building the PatchEmbedding layer

[what-is-available-for-name]

When I step-by-step debugged with the config (ViTPose_small_simple_coco_256x192.py),
I found a confusing value in the parameters of the PatchEmbedding layer as the padding below:
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=(patch_size[0] // ratio), padding=4 + 2 * (ratio//2-1))

When I run the train.py, the padding value would be padding = 4 + 2 * (1 // 2 - 1) = 2. It would cause the backbone to miss two columns at the right side of the origin image and two rows at the bottom of the origin image, because Patch_size = 16 and it divides evenly into the img_size(256x192).

Could you please explain the details about this issue?

[what-is-available-for-name]

1.pdf
A figure in pdf to help understand my issue