wtupc96 / channel-fusion

The author's officially PyTorch Channel-Fusion implementation.

Channel Fusion: An Effective Data Augmentation to Boost Image Classification

The author's officially PyTorch Channel-Fusion implementation.

The code of models is modified from kuangliu/pytorch-cifar, have the following modifications:

1. Support svhn dataset;
2. num_classes can be passed as a parameter;

• Thanks for the brilliant code of various models! 😄

Channel Fusion

Channel fusion is a surprisingly easy but effective data augmentation approach for models that only needs regrouping different image channels of the same class.

Requirements

• Python 3.6
• PyTorch 1.0

Run

1. Dataset preparation, to see data/README.md;
2. Train with python main.py --model [MODEL] --dataset [DATASET] (--channel_fusion) --lr [INIT_LR]
   • Available models:
     • VGG: VGG-11, VGG-13, VGG-16, VGG-19;
     • ResNet: ResNet-18, ResNet-34, ResNet-50, ResNet-101, ResNet-152;
     • ResNet-PreAct: ResNet-18-PreAct, ResNet-34-PreAct, ResNet-50-PreAct, ResNet-101-PreAct, ResNet-152-PreAct;
     • LeNet, GoogLeNet;
     • DenseNet: DenseNet-121, DenseNet-161, DenseNet-169, DenseNet-201;
     • ResNeXt: ResNeXt29_2x64d, ResNeXt29_4x64d, ResNeXt29_8x64d, ResNeXt29_32x4d;
     • MobileNet, MobileNet-V2;
     • DPN: DPN-26, DPN-92;
     • ShuffleNet: ShuffleNetG2, ShuffleNetG3, ShuffleNetV2;
     • SE-ResNet-18;
     • EfficientNet-B0.
   • Available datasets:
     • CIFAR-10;
     • CIFAR-100;
     • SVHN.
3. You will see results on the console and get checkpoints as well as losses and accuracies in checkpoint folder. :)

Results

ResNet-18-PreAct:

Training Loss	Training Error
Testing Loss	Testing Error