Progressive-GAN-pytorch

A pytorch implementation of Progressive-GAN that is actually work, readable and simple to customize

Description

I simplify the code of training a Progressive-GAN, making it easier to read and customize, for the purpose of research.
This implementation is portable with minimal library dependency (only torch and torchvision) and just 2 code modules. In the code, you can easily modeify the training-schema, the loss function, and the network structure, etc.
The key contributions in the paper: 1. progressively growing of GAN, 2. minibatch std on Discriminator, 3. pixel-norm on Generator, 4. equalized learning rate; are all implemented.
Enjoy the benefit of the progressive-growing infrastructure and port it to your own research and product!

How to run

To start a training, just run:

python train.py --path /path/to/image-folder

An example with more configuration can be:

python train.py --path /path/to/imagefolder --trial_name experiment-1 --z_dim 100 --channel 512 --batch_size 4 --init_step 2 --total_iter 300000 --pixel_norm --tanh

For a comprehensive explanation of all the parameters, run:

python train.py --help

Each new running of the code will create a new folder with the specified trail_name, all the generated images, model checkpoints and loss value loging file will be stored in this new folder. A copy of the codes that you run will also be intimately stored (because you might have modefied them).

Dataset

This code is ready for your own image datasets with the torchvision.datasets.ImageFolder module.
Place all your images in a way like:

<image_root_folder>
        |--<subfolder 1>
                |--image 1
                |--image 2 ...
        |--<subfolder 2>
        ...

Training results

This code performs consistently well on various datasets I tested, I just don't bother upload them here.

Reference