GAN-boilerplate

A boilerplate repo for GAN projects. Fork the repo when adapting for your own projects.

This repo is geared towards GANs for material microstructure projects, where the generator learned to output a homogeneous image. However, it can be easily adapted to any GAN project.

Folder structure

GAN-boilerplate
 ┣ src
 ┃ ┣ __init__.py
 ┃ ┣ networks.py
 ┃ ┣ postprocessing.py
 ┃ ┣ preprocessing.py
 ┃ ┣ test.py
 ┃ ┣ train.py
 ┃ ┗ util.py
 ┣ data
 ┃ ┗ example.png
 ┣ .gitignore
 ┣ config.py
 ┣ main.py
 ┣ README.md
 ┗ requirements.txt

Quickstart

For the repository.

Prerequisites:

• conda
• python3

Create a new conda environment, activate and install pytorch

Note: cudatoolkit version and pytorch install depends on system, see PyTorch install for more info.

conda create --name gan-boilerplate
conda activate gan-boilerplate
conda install pytorch torchvision -c pytorch
conda install -r requirements.txt

Create a .env file to hold secrets, the .env file must include

WANDB_API_KEY=
WANDB_ENTITY=
WANDB_PROJECT=

You are now ready to run the repo. To start training

python main.py train -t test-run

This will track your run online with Weights and Biases and name your training run test-run. To run in offline mode

python main.py train -t test-run -o

To generate samples from a trained generator

python main.py generate -t test-run

To run unit tests

python main.py test

Saving, loading and overwriting models

Models are saved to runs folder which is generated when training initiates. Inside runs, a new folder with the name of your training run tag will be generated, inside this the model params and training outputs are saved. This includes:

• config.json - this json holds the config paramaters of your training run, see config.py for more info
• Gen.pt - this holds the generator training parameters
• Disc.pt - this holds the discriminator parameters

Training

If training for the first time, these files are created and updated during training.

If you initiate a training run with a tag of a run that already exists you will see the prompt

To overwrite existing model enter 'o', to load existing model enter 'l' or to cancel enter 'c'.

By entering 'o' you will overwrite the existing models, deleting their saved parameters and config. l will load the existing model params and config, and continue training this model. c will abort the training run.

Evaluation

When evaluating a trained model, the params and model config are loaded from files. Models are saved with their training tag, use this tag to evaluate specific models.

TODO

• Quickstart
• Saving and loading models
• Training outputs
• Network architectures
• wandb