Reproduce the accuracy in paper about as-mlp HOT 4 CLOSED

@Gaffey Hi, as mentioned in our paper, the shift size=5, but the shift size=3 in the previous config file. We have modified the config file https://github.com/svip-lab/AS-MLP/blob/main/configs/as_tiny_patch4_shift5_224.yaml. Please check 4cb6e44. Using the new config, we reimplement the uploaded code and get 81.33 acc.

from as-mlp.

@dongzelian Thank you for your kind reply. Besides, I find a small difference between the code and the paper. The initial learning rate is set as 0.0005 in the code (BASE_LR in config.py) and 0.001 in the paper, which is the final used one?

from as-mlp.

@Gaffey Hi, the initial learning rate is 0.001, which is consistent with the value in the paper. Please note the batch size is 1024 (128/per GPU * 8GPUs) and refer to this line

from as-mlp.

@dongzelian got it, thanks again.

from as-mlp.