AugShuffleNet-plus

This repo presents a faster version of AugShuffleNet. We remove channel shuffle operation widely used in ShuffleNetV2 and AugShuffleNet, offering obvious improvement over training/inference speed.

python main.py --model augshufflenet

Results

All models are trained 200 epochs on cifar10 and cifar100. device: 3060 GPU.

10 runs on cifar10

Model	MAdds	Params	Acc(%)	training time/epoch
AugShuffleNet 1.5x	85.38M	2.22M	94.00 93.95 94.16 93.98 93.86 94.07 93.94 93.83 93.45 94.12	21.4s
ShuffleNetV2 1.5x	94.27M	2.49M	93.92 93.47 93.28 93.59 93.67 93.71 93.78 93.55 93.44 93.72	26.2s
AugShuffleNet 1.0x	43.55M	1.21M	93.32 93.51 93.38 93.50 93.39 93.50 93.43 93.13 93.27 93.44	14.7s
ShuffleNetV2 1.0x	45.01M	1.26M	92.88 92.98 92.75 92.93 92.87 92.87 93.14 92.69 92.80 93.05	17.5s
AugShuffleNet 0.5x	10.20M	0.33M	90.24 90.75 91.17 90.46 91.01 91.29 90.92 91.03 91.20 90.91	7.3s
ShuffleNetV2 0.5x	10.91M	0.35M	90.23 90.74 90.56 90.58 90.31 89.96 90.66 89.78 90.76 90.05	8.8s

10 runs on cifar100

Model	MAdds	Params	Acc	training time/epoch
AugShuffleNet 1.5x	85.5M	2.32M	74.12 74.55 74.95 74.88 74.78 74.27 74.90 74.61 74.36 74.68	21.4s
ShuffleNetV2 1.5x	94.36M	2.58M	73.47 73.68 73.64 73.76 74.10 73.74 73.58 73.87 73.35 73.35	26.2s
AugShuffleNet 1.0x	43.65M	1.30M	73.65 72.87 72.82 72.81 73.38 73.54 72.57 73.19 72.41 73.09	14.7s
ShuffleNetV2 1.0x	45.10M	1.36M	71.70 71.97 72.11 71.91 71.62 71.95 72.22 71.49 72.08 72.28	17.5s
AugShuffleNet 0.5x	10.29M	0.42M	68.95 68.67 68.90 68.28 68.48 68.20 68.22 69.45 68.04 68.06	7.3s
ShuffleNetV2 0.5x	11.00M	0.44M	67.39 66.70 67.47 66.12 66.55 67.44 66.74 66.82 66.52 66.76	8.8s

Citation

@article{ye2022augshufflenet,
  title={AugShuffleNet: Communicate more, compute less},
  author={Ye, Longqing},
  journal={arXiv preprint arXiv:2203.06589},
  year={2022}
}