https://github.com/wtupc96/channel-fusion

The author's officially PyTorch Channel-Fusion implementation.
https://github.com/wtupc96/channel-fusion

cifar10 cifar100 classification color-space computer-vision data-augmentation deep-learning image-processing svhn transformation

Last synced: 4 months ago
JSON representation

The author's officially PyTorch Channel-Fusion implementation.

• Host: GitHub
• URL: https://github.com/wtupc96/channel-fusion
• Owner: wtupc96
• License: cc0-1.0
• Created: 2019-12-25T09:20:56.000Z (almost 6 years ago)
• Default Branch: master
• Last Pushed: 2019-12-27T06:34:53.000Z (almost 6 years ago)
• Last Synced: 2025-02-25T09:24:00.686Z (7 months ago)
• Topics: cifar10, cifar100, classification, color-space, computer-vision, data-augmentation, deep-learning, image-processing, svhn, transformation
• Language: Python
• Size: 313 KB
• Stars: 3
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# 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`](https://github.com/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!* :smile:

## 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.

![](imgs/channel_fusion_downsample.png)

## Requirements

- Python 3.6

- PyTorch 1.0

## Run

1. Dataset preparation, to see [`data/README.md`](https://github.com/wtupc96/channel-fusion/tree/master/data);

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:

| ![](imgs/training_loss.png) | ![](imgs/training_error.png) |

| :-------------------------: | :--------------------------: |

|        Training Loss        |        Training Error        |

| ![](imgs/testing_loss.png)  | ![](imgs/testing_error.png)  |

|        Testing Loss         |        Testing Error         |