/fashion-mnist-cgan

CGAN and DCGAN for Fashion MNIST dataset

Primary LanguagePython

fashion-mnist-cgan

Architectures

  • Conditional Generative Adversarial Nets (2014)
  • Deep Convolutional Generative Adversarial Networks (2016)

Results

Interactive report: link

Generator output progress

Model Epoch 0 Epoch 12 Epoch 24
CGAN
DCGAN

Loss

Model BCE Loss
CGAN
DCGAN

Overall DCGAN achieved better results, not only in terms of generator loss, but also visually pleasing results. However, the training took twice as much time as CGAN training.

Running

  1. Build the container.
docker build -t fashion .
  1. Run.
docker run --runtime=nvidia fashion

If you want to use Weights & Biases make sure to assign your key to WANDB_API_KEY env variable and run:

docker run --runtime=nvidia --env-file .env fashion --wandb 1

Further improvements

  • Pretraining generator model
  • Testing other architectures
  • More extensive hyperparameter tuning (more parameters and more optuna trials)
  • Writing tests