ShuffleNet in PyTorch

An implementation of ShuffleNet in PyTorch. ShuffleNet is an efficient convolutional neural network architecture for mobile devices. According to the paper, it outperforms Google's MobileNet by a small percentage.

What is ShuffleNet?

In one sentence, ShuffleNet is a ResNet-like model that uses residual blocks (called ShuffleUnits), with the main innovation being the use of pointwise, or 1x1, group convolutions as opposed to normal pointwise convolutions.

Usage

Clone the repo:

git clone https://github.com/jaxony/ShuffleNet.git

Use the model defined in model.py:

from model import ShuffleNet

# running on MNIST
net = ShuffleNet(num_classes=10, in_channels=1)

Performance

Trained on ImageNet (using the PyTorch ImageNet example) with groups=3 and no channel multiplier. On the test set, got 62.2% top 1 and 84.2% top 5. Unfortunately, this isn't comparable to Table 5 of the paper, because they don't run a network with these settings, but it is somewhere between the network with groups=3 and half the number of channels (42.8% top 1) and the network with the same number of channels but groups=8 (32.4% top 1). The pretrained state dictionary can be found here, in the following format:

{
    'epoch': epoch + 1,
    'arch': args.arch,
    'state_dict': model.state_dict(),
    'best_prec1': best_prec1,
    'optimizer' : optimizer.state_dict()
}

Note: trained with the default ImageNet settings, which are actually different from the training regime described in the paper. Pending running again with those settings (and groups=8).