NITR098/AttSwinUNet

from utils import load_pretrain

Closed this issue · 3 comments

ZhouHongLiang6 commented

I can‘t find the utils files，when I run the rode，it tell me it can't find the load_pretrain,can you help me?

TWENTY-FOU commented

同问
解决了可以艾特一下我嘛？

Kahsolt commented

That load_pretrain seems redundant and can be just safely removed, see line 104~106 for the real load of pretrained weights:

if int(config['pretrained']):
    Net.load_state_dict(torch.load(config['saved_model'], map_location='cpu')['model_weights'])
    best_val_loss = torch.load(config['saved_model'], map_location='cpu')['val_loss']

And also, when it comes to error missing xcit_tiny_12_p16, just comment that two lines, they are NOT EVEN USED.
And then, when trainer is a module and cannot be called, modify its import clause from trainer import trainer.
And finally, I got AssertionError: Input image size (256*256) doesn't match model (224*224). only to find that the pretrained model is sized 244 and cannot even apply on ANY DATASET PRESETS sized in 256 as README.md declares!!!!!

After all, I DO NOT really know why this repo is soooo broken...
Is this really an official implementation for a published or ready-to-publish work??
It upsets me a bit :(

  • 👀1