gettingstarted.md

Getting started

HyperGAN is currently in open beta.

Everyone will have different goals when using hypergan. Here are some common uses:

Training a network

HyperGAN is currently beta. We are still searching for a default cross data-set configuration. Various papers are implemented by listing `hypergan new modelname -l'

Deploying a model

Building datasets

Building image datasets

Using search

Each of the examples support search. Automated search can help find good configurations.

If you are unsure, you can start with the 2d-distribution.py. Check out random_search.py for possibilities, you'll likely want to modify it.

Please consider sharing in a PR if you find goood configurations!

Manual search

It's faster to do manual search if you are actively looking at results. For manual search here are some tips:

• create a baseline json configuration
• start with a working configuration
• run baseline every time
• each phase make the best performer the new baseline
• the examples are capable of (sometimes) finding a good trainer, like 2d-distribution. Mixing and matching components seems to work.

Creating a new configuration

Hyperparameter tuning, or doing something like trying softmax loss with gradient penalty, can all be done in the json configuration file.

Implementing a paper

Each paper is a combination of json file and code.

HyperGAN has tried to make it easy to add a new component. Here are some basic classification rules on where paper implementations should go:

paper type	proposed implementation
new type of encoder	new encoder class
new type of input distribution encoding	either new encoder class or projection of uniform_encoder
new type of generator	create a new generator class
feature that could possibly apply to all generators	BaseGenerator
new type of discriminator	create a new discriminator class
applies to all discriminators	add it to all discriminators with BaseDiscriminator
new type of loss	create a new loss class
new type of gradient penalty	add it to all losses with BaseLoss
new gan design	new GAN
new trainer	configuration change
new activation	add it to ops.py. Be sure to include it in the huge conditional, search for prelu and you'll see
something else	Probably easiest to hack it in as a GAN, even if you have to build the graph up yourself in #create.

When implementing a new paper, feature-gate any breaking code so that old configurations still work(and you have something to compare).

Custom research

HyperGAN is meant to support custom research as well. You can replace any part of the GAN with the json file, or just create a new GAN altogether.

Also, try to save refactoring until after it works.

Using hyperGAN in an app

hypergan build

This builds a standard tensorflow graph. You can then quantize and/or deploy the model.