https://github.com/jaxony/ShuffleNet

ShuffleNet in PyTorch. Based on https://arxiv.org/abs/1707.01083
https://github.com/jaxony/ShuffleNet

artificial-intelligence convolution deep-learning neural-network pytorch

Last synced: about 1 month ago
JSON representation

ShuffleNet in PyTorch. Based on https://arxiv.org/abs/1707.01083

• Host: GitHub
• URL: https://github.com/jaxony/ShuffleNet
• Owner: jaxony
• License: mit
• Created: 2017-07-15T06:58:13.000Z (over 7 years ago)
• Default Branch: master
• Last Pushed: 2017-12-20T12:50:24.000Z (almost 7 years ago)
• Last Synced: 2024-08-01T22:50:06.003Z (4 months ago)
• Topics: artificial-intelligence, convolution, deep-learning, neural-network, pytorch
• Language: Python
• Homepage:
• Size: 13.4 MB
• Stars: 290
• Watchers: 10
• Forks: 90
• Open Issues: 6
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

• awesome-image-classification - unofficial-pytorch : https://github.com/jaxony/ShuffleNet
• awesome-image-classification - unofficial-pytorch : https://github.com/jaxony/ShuffleNet

README

        
# 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:

```bash

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

```

Use the model defined in `model.py`:

```python

from model import ShuffleNet

# running on MNIST

net = ShuffleNet(num_classes=10, in_channels=1)

```

## Performance

Trained on ImageNet (using the [PyTorch ImageNet example][imagenet]) 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][tar], in

the [following

format](://github.com/pytorch/examples/blob/master/imagenet/main.py#L165-L171):

```

{

    '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`).

[tar]: https://drive.google.com/file/d/12oGJsyDgp51LhQ7FOzKxF9nBsutLkE6V/view?usp=sharing

[imagenet]: https://github.com/pytorch/examples/tree/master/imagenet