cutmix

a Ready-to-use PyTorch Extension of Unofficial CutMix Implementations.

This re-implementation is improved in some parts,

• Fixing issue #1 in the original repository
• issue #3 : Random crop regions are randomly chosen, even within the same batch.
• issue #4 : Different lambda values(sizes of crop regions) are randomly chosen, even within the same batch.
• Images to be cropped are randomly chosen in the whole dataset. Original implementation selects images only inside the same batch(shuffling).
• Easy to install and use on your existing project.
• With additional augmentations(fast-autoaugment), the performances are improved further.

Hence, there may be slightly-improved training results also.

Requirements

• python3
• torch >= 1.1.0

Install

This repository is pip-installable,

$ pip install git+https://github.com/ildoonet/cutmix

or you can copy 'cutmix' folder to your project to use it.

Usage

Our CutMix is inhereted from the PyTorch Dataset class so you can wrap your own dataset(eg. cifar10, imagenet, ...). Also we provide CutMixCrossEntropyLoss, soft version of cross-entropy loss, which accept soft-labels required by cutmix.

from cutmix.cutmix import CutMix
from cutmix.utils import CutMixCrossEntropyLoss
...

dataset = datasets.CIFAR100(args.cifarpath, train=True, download=True, transform=transform_train)
dataset = CutMix(dataset, num_class=100, beta=1.0, prob=0.5, num_mix=2)    # this is paper's original setting for cifar.
...

criterion = CutMixCrossEntropyLoss(True)
for _ in range(num_epoch):
    for input, target in loader:    # input is cutmixed image's normalized tensor and target is soft-label which made by mixing 2 or more labels.
        output = model(input)
        loss = criterion(output, target)
    
        loss.backward()
        optimizer.step()
        optimizer.zero_grad()

Result

PyramidNet-200 + ShakeDrop + CutMix \w CIFAR-100

	Top-1 Error(@300epoch)	Top-1 Error(Best)	Model File
Paper's Reported Result	N/A	13.81	N/A
Our Re-implementation	13.68	13.15	Download(12.88)
+ Fast AutoAugment	13.3	12.95

We ran 6 indenpendent experiments with our re-implemented codes and got top-1 errors of 13.09, 13.29, 13.27, 13.24, 13.15 and 12.88, using below command. (Converged at 300epoch with the top-1 errors of 13.55, 13.66, 13.95, 13.9, 13.8 and 13.32.)

$ python train.py -c conf/cifar100_pyramid200.yaml

ResNet + CutMix \w ImageNet

		Top-1 Error (@300epoch)	Top-1 Error (Best)	Model File
ResNet18	Reported Result \wo CutMix	N/A	30.43
	Ours	29.674	29.56
ResNet34	Reported Result \wo CutMix	N/A	26.456
	Ours	24.7	24.57	Download
ResNet50	Paper's Reported Result	N/A	21.4	N/A
	Author's Code(Our Re-run)	21.768	21.586	N/A
	Our Re-implementation	21.524	21.340	Download(21.25)
ResNet200	Our Re-implementation
	+ Fast AutoAugment	19.058	18.858

$ python train.py -c conf/imagenet_resnet50.yaml

We ran 5 independent experiments on ResNet50.

• Author's codes

  • 300epoch : 21.762, 21.614, 21.762, 21.644, 21.810
  • best : 21.56, 21.556, 21.666, 21.498, 21.648

• Our Re-implementation

  • 300epoch : 21.53, 21.408, 21.55, 21.4, 21.73
  • best : 21.392, 21.328, 21.386, 21.256, 21.34

Reference

• Official
  • Paper : https://arxiv.org/abs/1905.04899
  • Implementation : https://github.com/clovaai/CutMix-PyTorch
• ShakeDrop
  • https://github.com/owruby/shake-drop_pytorch
• Fast AutoAugment
  • https://github.com/kakaobrain/fast-autoaugment