PyTorch Implementation of shake-shake

Usage

$ python train.py --depth 26 --base_channels 32 --shake_forward True --shake_backward True --shake_image True --outdir results

Results on CIFAR-10

Model	Test Error (median of 3 runs)	Test Error (in paper)	Training Time
shake-shake-26 2x32d (S-S-I)	3.68	3.55 (average of 3 runs)	33h49m
shake-shake-26 2x64d (S-S-I)	2.88 (1 run)	2.98 (average of 3 runs)	78h48m
shake-shake-26 2x96d (S-S-I)	2.90 (1 run)	2.86 (average of 5 runs)	101h32m*

Notes

• The model of shake-shake-26 2x64d (S-S-I) is trained with batch size 64, and initial learning rate 0.1.

References

• Gastaldi, Xavier. "Shake-Shake regularization." In International Conference on Learning Representations, 2017. arXiv:1705.07485, Torch implementation