https://github.com/neuralix/google_evolution

https://github.com/neuralix/google_evolution

Last synced: about 2 months ago
JSON representation

• Host: GitHub
• URL: https://github.com/neuralix/google_evolution
• Owner: neuralix
• Created: 2017-03-22T10:23:59.000Z (over 7 years ago)
• Default Branch: master
• Last Pushed: 2017-03-23T04:46:45.000Z (over 7 years ago)
• Last Synced: 2024-04-20T10:35:26.569Z (5 months ago)
• Language: Python
• Size: 70.3 KB
• Stars: 53
• Watchers: 4
• Forks: 15
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
Exotic structured image classifier

=====================================

This implements one of result networks from [Large-scale evolution of image classifiers](https://arxiv.org/abs/1703.01041) by Esteban Real, et. al.

## Requirements

- Install [pytorch](http://pytorch.org/) (I recommend anaconda environment.)

- Install [scikit-learn](http://scikit-learn.org/stable/)

- Download [CIFAR10 dataset](https://www.cs.toronto.edu/~kriz/cifar.html)

## Training

1. Copy two files to `{torchvision_path}/models`

```bash

cp {__init__.py,evloution.py} {torchvision_path}/models

```

2. Run

```bash

python main.py -a evolution {cifar10_data_dir}

```

cf. How to know {torchvision_path}?

```

import torchvision

print(torchvision.__file__)

```

## Note

1. The numbers of channels are not in the paper and it is set by me similar with vgg.

You need to adjust these for better performance.

2. If you want to adsjust learning rate on-the-fly, create `lr.txt` having lr value in same directory with main.py. And just change the value before some epoch you want to adopt new lr value.

3. Used CIFAR10 is image file dataset. It's not CIFAR-10 {python,Matlab,binary} version.

You should convert it to conventional image files. It you don't want it, you have to use your proper dataloader.

## Implemented network