pytorch_cutmix's Introduction

PyTorch Implementation of CutMix

Usage

$ python train.py --depth 20 --use_cutmix --outdir results

Results on CIFAR-10

Model	Test Error (median of 3 runs)	Training Time
WRN-20-4	4.56	1h22m
WRN-20-4, CutMix (alpha=1)	3.62	1h22m

These models were trained for 300 epochs with batch size 128, initial learning rate 0.2, and cosine annealing.
Test errors reported above are of the last epoch.
These experiments were done using Tesla V100.

w/o CutMix

$ python -u train.py --depth 20 --base_channels 64 --base_lr 0.2 --scheduler cosine --seed 7 --outdir results/wo_cutmix/00

w/ CutMix

$ python -u train.py --depth 20 --base_channels 64 --base_lr 0.2 --scheduler cosine --seed 7 --use_cutmix --cutmix_alpha 1.0 --outdir results/w_cutmix/00

References

Yun, Sangdoo, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. "CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features." arXiv preprint arXiv:1905.04899 (2019). arXiv:1905.04899

pytorch_cutmix's People

Contributors

Stargazers

Watchers

pytorch_cutmix's Issues

argument 'tensors' (position 1) must be tuple of Tensors, not Tensor

`
dataset == 'imagenet':
normalize = T.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
collator = CutMixCollator(alpha=1.0)

    train_transform = T.Compose([
        T.RandomResizedCrop(64),
        T.RandomHorizontalFlip(),
        T.ToTensor(),
        collator,
        normalize,
        ]) 
    
    test_transform = T.Compose([
        T.ToTensor(),               #(224, 224) /imagenetFromServer/train
        T.CenterCrop((64)),    #(224, 224) for toy_dataset/train   # (32, 32) is for following folder ./data/imagenet/train
        T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) # T.Normalize((0.5071, 0.4867, 0.4408), (0.2675, 0.2565, 0.2761)) # CIFAR-100
    ])

`
/data/_utils/collate.py", line 56, in default_collate
return torch.stack(batch, 0, out=out)
TypeError: stack(): argument 'tensors' (position 1) must be tuple of Tensors, not Tensor

Accuracy Graph

Nice work man.
Can you please upload the code for drawing graph of accuraccy?

Thank you very much for your open source code. Recently, I have implemented your code, but I am unable to obtain the image processed by the code. I would like to ask how to save the image enhanced image. I really hope to receive your reply. Thank you

Recommend Projects