Potential mistake in SegFormer model: `patch_size` argument in SegFormer model not being used.

[jonasdieker]

Hi there,

first of all thank you for your work and providing all the code! I was looking at the following lines in the SegFormer backbone model:

SegFormer/mmseg/models/backbones/mix_transformer.py

Lines 203 to 220 in 65fa8cf

	class MixVisionTransformer(nn.Module):
	def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dims=[64, 128, 256, 512],
	num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0.,
	attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm,
	depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1]):
	super().__init__()
	self.num_classes = num_classes
	self.depths = depths
	# patch_embed
	self.patch_embed1 = OverlapPatchEmbed(img_size=img_size, patch_size=7, stride=4, in_chans=in_chans,
	embed_dim=embed_dims[0])
	self.patch_embed2 = OverlapPatchEmbed(img_size=img_size // 4, patch_size=3, stride=2, in_chans=embed_dims[0],
	embed_dim=embed_dims[1])
	self.patch_embed3 = OverlapPatchEmbed(img_size=img_size // 8, patch_size=3, stride=2, in_chans=embed_dims[1],
	embed_dim=embed_dims[2])
	self.patch_embed4 = OverlapPatchEmbed(img_size=img_size // 16, patch_size=3, stride=2, in_chans=embed_dims[2],
	embed_dim=embed_dims[3])

I noticed that the argument patch_size is not actually being used for the OverlapPatchEmbed modules.

Instead you hard coded a patch sizes of [7, 3, 3, 3] for the 4 blocks. While this of course is still smaller than the 16x16 patches in ViT, and thus still lends itself better to detection and segmentation tasks, the model deviates from the paper, where you describe an initial patch size of 4 being used. This also means that classes inheriting from this class do not use the argument at all!

Maybe I am misunderstanding something, so I would be happy if you could shed some light on this potential mistake! Thank you.

[hubert10]

Hi Jonas, I've also observed something similar and this needs to be clarified either in the paper or in the code above!