GAN Implementation in PyTorch

This repository implements training and inference methods of GAN with just fc layers on MNIST.

GAN Explanation & Implementation Video

Output on MNIST

Data preparation

For setting up the mnist dataset:

Follow - https://github.com/explainingai-code/Pytorch-VAE#data-preparation

The directory structure should look like this

$REPO_ROOT
    -> data
        -> train
            -> images
                -> 0
                    *.png
                -> 1
                ...
                -> 9
                    *.png
        -> test
            -> images
                -> 0
                    *.png
                ...
    -> dataset
    -> tools
        

Quickstart

• Create a new conda environment with python 3.8 then run below commands
• git clone https://github.com/explainingai-code/GANs-Pytorch.git
• cd GANs-Pytorch
• pip install -r requirements.txt
• python -m tools.train_gan for training and saving inference samples

Training on colored mnist images

• Ensure dataset is prepared according to the Data Preparation instructions
• Change the IM_CHANNELS field to 3 in train_gan.py
• Uncomment lines 56-59 in the dataset/mnist_dataset.py file

Training on custom dataset images

• Dump all *.png files(or whatever format images you have) in the path data/train/images
• Comment https://github.com/explainingai-code/GANs-Pytorch/blob/main/dataset/mnist_dataset.py#L43
• Directory structure should be following:

data
    -> train
        -> images 
            *.png           

• Change the IM_PATH field to data/train in train_gan.py
• Change the channels and image sizes accordingly

Output

Outputs will be saved every 50 steps in samples directory .

During training of GAN the following output will be saved

• Latest Model checkpoints for generator and discriminator in $REPO_ROOT directory

During inference every 50 steps the following output will be saved

• Sampled image grid for in samples/*.png

Citations

@misc{goodfellow2014generative,
      title={Generative Adversarial Networks}, 
      author={Ian J. Goodfellow and Jean Pouget-Abadie and Mehdi Mirza and Bing Xu and David Warde-Farley and Sherjil Ozair and Aaron Courville and Yoshua Bengio},
      year={2014},
      eprint={1406.2661},
      archivePrefix={arXiv},
      primaryClass={stat.ML}
}