FcaNet-CIFAR

An implementation of the paper FcaNet: Frequency Channel Attention Networks on CIFAR10/CIFAR100 dataset.

how to run

Code: python Cifar.py --network fca_resnet20

Updates

2020-12-31 Conducted some experiments on YOLOv3 and got some positive results. Code will be uploaded in few days.

Notes

• This project is only for my own study purpose. Please don't star this project because I'm not one of the authors. If you want to try FcaNet, welcome to use the codes and follow the author of the paper--cfzd.
• The basic code architecture is based on SENet-cifar10. Very few tricks are utilized, so the performance may not be satisfying.
• The CIFAR datasets are pretty small compared with ImageNet, so the experiments are not stable and representative for verifying the algorithm. More experiments on ImageNet will coming soon.

Experiment

Denpendencies

• python 3.7
• pytorch 1.4.0
• torchvision 0.5.0

Conditions

• Data augmentation: pad=4, crop=32; horizontal flip
• optim: default = SGD(lr=0.1,m=0.9,wd=1e-4, bs=128)

Experiments with different network archs and attentions.

2020-12-28 update: Fix the DCT weights indexes calculation. Also referring to author's comment in zhihu

Ablation Study about dct_weights

refer to the comments in zhihu.