input image=112*112

[eeric]

@volgachen
can you provide model structure with input=112*112,?
thanks!

[volgachen]

There are several ways to modify the original model structure. For example,

• Change argument patch_size from 4 to 2 to modify the patch embedding module at the very beginning.

model = PyramidVisionTransformer(
    patch_size=2, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], qkv_bias=True,
    norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1], **kwargs)

• Remove a certain stage (I prefer the second stage) and its corresponding DePatch module.
• Simply run the original model without any modification. (Setting img_size=112 at here).

However, I cannot tell which could be the best without experiments. You can try it. I am glad to see your results.

[eeric]

thanks!

1.PyramidVisionTransformer, failed;
2.Remove a certain stage, complicated, failed;
3.Setting img_size=112, failed.