Implement Vanilla GAN architecture using PyTorch.
- The generator and discriminator were implemented in PyTorch, wherein each use 4 linear layers.
- The GAN architecture was trained on MNIST handwriting dataset.
- The generator is fed with 128 dimension noise input and it generates 28x28 grayscale handwriting data consisting of digits.
- To train the model with the default configurations,
run the command:
python3 run.py - To change the epochs, batch size, patience, run the command with necessary
arguments.
Example:python3 --epochs=100 --batch_size=128 patience=10 - After training, the generator model and discriminator model with the best loss will be saved at ./checkpoints/generator.pt and ./checkpoints/discriminator.pt respectively.
- Sample generated images and loss plots of training are stored in the folder ./results/
- The model converged around 100 epochs, generating adeqaute images resembling the original dataset.
- The loss plot during the training phase is as follows:
- Generation of images over the epochs can be visualized as:
