The problem is shown in the figure

[rui-cf]

model = CSWinTransformer(patch_size=4, embed_dim=96, depth=[2,4,32,2], split_size=[1,2,12,12], num_heads=[4,8,16,32], mlp_ratio=4.).cuda().eval() inp = torch.rand(1, 3, 224, 224).cuda() outs = model(inp) for out in outs: print(out.shape)

RuntimeError: shape '[1, 192, 1, 14, 1, 12]' is invalid for input of size 37632

why?

[LightDXY]

The feature map size should be divisible by the split size. For example, with image_size=224, the feature map size of stage 3 is 14, and the default split size is 7. We use split size=12 for input image size=384.

We add padding to realize a flexible split size for downstream tasks, but we do not consider it in the ImageNet model because the padding is not efficient.

[506813254]

The feature map size should be divisible by the split size. For example, with image_size=224, the feature map size of stage 3 is 14, and the default split size is 7. We use split size=12 for input image size=384.

We add padding to realize a flexible split size for downstream tasks, but we do not consider it in the ImageNet model because the padding is not efficient.

Hi，What does the padding do？

[54wb]

The feature map size should be divisible by the split size. For example, with image_size=224, the feature map size of stage 3 is 14, and the default split size is 7. We use split size=12 for input image size=384. We add padding to realize a flexible split size for downstream tasks, but we do not consider it in the ImageNet model because the padding is not efficient.

Hi，What does the padding do？

hi, do you know what does the padding do or how can i find the padding code

[wujiang0156]

the table's list is as same as the code. which is right? thank you