Shake-Shake regularization

PyTorch implementation of shake-shake regularization. Author implementations is here.

Dependencies

• python 3.5
• PyTorch 0.4.1

Accuracy

CIFAR-10

Model	Method	This implementaion	Paper
ResNet26-2x32	S-S-I	96.37	96.45
ResNet26-2x64	S-S-I	97.07	97.02

CIFAR-100

Model	Method	This implementaion	Paper
ResNeXt29-2x4x64d	S-S-I	TODO	84.03

Train ResNet26-2x64d for CIFAR-10

python train.py --label 10 --depth 26 --w_base 64 --lr 0.1 --epochs 1800 --batch_size 64

Effects of learning epoch on accuracy

Model	Method	Epoch	This implementaion	Paper
ResNet26-2x64	S-S-I	300	96.46	-
ResNet26-2x64	S-S-I	600	96.65	-
ResNet26-2x64	S-S-I	900	96.78	-
ResNet26-2x64	S-S-I	1200	96.68	-
ResNet26-2x64	S-S-I	1500	96.77	-
ResNet26-2x64	S-S-I	1800	97.07	97.02

References

Gastaldi, Xavier. "Shake-Shake regularization." arXiv:1705.07485v2

Gastaldi, Xavier. "Shake-Shake regularization of 3-branch residual networks." ICLR2017 Workshop