ggjy/hire-wave-mlp.pytorch

Implementation of Hire-MLP: Vision MLP via Hierarchical Rearrangement and An Image Patch is a Wave: Phase-Aware Vision MLP.

hire-wave-mlp.pytorch's Issues

I am looking forward to the application of this method to the field of semantic segmentation, and hope to make the code public as soon as possible

Difference of PATM between code and paper

In Section 3.2 of the paper, the authors say that "The final output of the block is the summation of these three branches." However, it seems that there are extra reweight and projection after the summation. Maybe this design is from CycleMLP, but not mentioned in the paper, which may cause confusion.

Single GPU WaveMLPT-dw training.

Hello,
I'm trying to run your repo with a small batch size of 2 on my modest RTX2080Su. A simple single run. No parallelism. I understand it is impractical to train the whole thing here. But I just want it to run for 1 epoch -- a small sanity check.

But trying to run the file on my machine, I receive the following 2 errors (2nd one of which is blocking):

1. File train_wave.mlp On line 578: args does not have max_history. Perhaps this is an argument generated on author's machine after the first run. I choose to remove this argument from CheckpointSaver function. This seems to go through.

2. File train_wave.mlp In function train_epoch On line 668: for some reason which I don't understand, the code passes in target.argmax(dims = -1) to the accuracy() function from timm.utils, instead of directly feeding it target. This causes an error. I instead just pass in target and the code works. Note, that contrastingly, in function validate on line 780 the code chooses to pass target and not the quirky argmax version.

It would really help to get some clarity on this peculiar behavior.