A versatile GAN(generative adversarial network) implementation focused on scalability and ease-of-use.

The following are a small sample of the manifolds that our generators have learned.

Randomly chosen:

• Card game

• People

• Fonts (the letter 'r')

• Efficient GAN implementation
• Semi-supervised or unsupervised learning(works with and without labels)
• Variational methods
• InfoGAN-inspired categories
• Minibatch normalization
• Adversarial inference
• Flask server mode

• --directory, required, specifies which parent directory to use. Each subdirectory is a different classification. For example if you have 'data/a' and 'data/b', files in 'data/a' will be represented internally as the one-hot vector [1,0].
• --channels, optional(default 3), the number of channels in your images. Black and white images typically only have 1.
• --width, optional(default 64), the width of each image. If smaller than this width, the image will be centered and cropped.
• --height, optional(default 64), the height of each image. If smaller than this height, the image will be centered and cropped.
• --format, optional(default png), file format of the images. Only supports jpg and png for now.
• --epoch, optional(default 10), number of epochs to run before exiting
• --load_config, optional(default None), the config uuid from hyperchamber to run
• --save_every, optional(default 0), after this many epochs, the network weights/checkpoint are saved into the 'saves' directory.
• --server, optional(default False), this will turn the app into server mode. Currently undocumented.

First, choose a dataset. If you use one of the standard datasets, you can find a config that will work with hyperchamber. If using a new/different dataset, start with a 'Search', detailed below.

For example, the following command will run a configuration with id https://hyperchamber.255bits.com/ba3051e944601e2d98b601f3347df0b1/40k_overfit_3:1.2/samples/25e15455541d84bb76dcc4c8f8dce5a1

Note: Requires a (free) account at hyperchamber to run.

  python3 directory-gan.py --directory test/ --channels 3 --width 64 --height 80 --format jpg --epoch 10000 --load_config 6d09d839d3b74a208a168dc3189c3f59 --save_every 10

  python3 directory-gan.py --directory test/ --channels 3 --width 64 --height 80 --format jpg --epoch 1
```

You can review the results on hyperchamber.

## How to use

Fair warning, this is all still alpha.  Even so, email us if you have issues running this on your dataset.

To use on any data:

* Run config sweep with hyperchamber
* Run your favorite config for longer


## Papers

* GAN - https://arxiv.org/abs/1406.2661
* DCGAN - https://arxiv.org/abs/1511.06434
* InfoGAN - https://arxiv.org/abs/1606.03657
* Improved GAN - https://arxiv.org/abs/1606.03498
* Adversarial Inference - https://arxiv.org/abs/1606.00704

## Sources

* DCGAN - https://github.com/carpedm20/DCGAN-tensorflow
* InfoGAN - https://github.com/openai/InfoGAN
* Improved GAN - https://github.com/openai/improved-gan
* Hyperchamber - https://github.com/255bits/hyperchamber