GAN

Python implementation of GAN and DCGAN using pytorch
Based on the following papers:
GAN - https://arxiv.org/pdf/1406.2661.pdf
DCGAN - https://arxiv.org/pdf/1511.06434.pdf

The results from training for 30 epochs is as follows:

The original hyperparameters were:

• Epochs= 30
• Input noise size = 100
• Batch size = 256
• Optimizer = Adam
  • lr=2e-4, betas=(0.5, 0.999) for Generator
  • lr=5e-4, betas=(0.5, 0.999) for Discriminator
• Loss function = BCE loss

About the repo

• main.py contains the training loop for the GAN
• Notes_GAN.md contains notes and algorithm from the papers
• remember to check the DATA_PATH variable, and change it to your local directory
• there is a Generator and Discriminator for both Multilayer perceptron model as well as CNN
• you can train the GAN using either, all you have to do is use Generator or GeneratorConv likewise Discriminator or DiscriminatorConv as generator and discriminator in the training loop
• the training loop will generate the saved models and checkpoints, which you will need for generating gifs and images
• viewOutput.py contains the code for generating gifs or images of the output of the GAN
• to generate gifs use the showAnimation() function which takes in the Generator net and device (i.e., cpu or gpu)
• to generate images use the showOutputGrid() function which takes in the Generator net and device (i.e., cpu or gpu)