shuffleNet-cifar10

a pytorch implement of shuffleNet on cifar-10

channel shuffle

• channel shuffle is a operation proposed in shuffleNet to adress the information isolation between channels while using successive group convolution.

• It can be done using only several lines code

  # channel shuffle
  n, c, w, h = x.shape
  x = x.view(n, self.g, self.n, w, h)
  x = x.transpose_(1, 2).contiguous()
  x = x.view(n, c, w, h)

• there is a demo picture to show what happened when channel shuffle

  

training on cifar10

• To make it suit cifar10's image size, I have disabled some downsample operation (i.e. maxpooling or stride = 2) and just keep the last two

• because of the low efficiency of group convolution, it takes relatively long time to train, more details can be seen below

  scale factor	groups	params/M	flops/M	training time	accuracy
  1.0	8	0.9131	161.70	11.4h	92.29%
  0.5	8	0.2507	43.43	6.5h	91.48%
  0.5	3	0.2427	42.97	4.0h	92.60%
  0.5	1	0.2487	44.63	3.6h	91.44%

  • here the accuracy means the max accuracy on validation set

  • training time is measured on a titan x (pascal) GPU

  • the results is comparable with resnet 20 which have the similar number of parameters:

    resnet 20 params: 0.27M accuracy: 91.25%

• more logs and the best weights can be get in folder bak

dependency

• pytorch 0.4.0
• python 3.x