facial-micro-expressiongc / dcgan

Keras implementation of Deep Convolutional Generative Adversarial Networks (DCGAN)

Keras implementation of Deep Convolutional Generative Adversarial Networks (DCGAN)

This is an implementation of DCGAN (Link to the paper: http://arxiv.org/abs/1511.06434) with Keras on top of TensorFlow. Two adversarial networks are trained on real images for generating artificial images that seems real.

Requirements

Install using pip install -r requirements.txt

You will also need to ensure the proper CUDA libraries and NVIDIA drivers are installed.

Usage

The dcgan.py script enables training of the DCGAN model with MNIST dataset and subsequently generate artificial images from the trained model.

Training:

python dcgan.py --mode train --batch_size <batch_size> --num_epoch <num_epoch>

Example:

python dcgan.py --mode train --batch_size 128 --num_epoch 100

Generate images:

python dcgan.py --mode generate --batch_size <batch_size>

Note: the batch_size value for generating images must equal to the batch_size value used during the training step.

Similarly, the optional --pretty flag will generate the top 5% artificial image determined by the discriminator.

python dcgan.py --mode generate --batch_size <batch_size> --pretty

Example:

python dcgan.py --mode generate --batch_size 128

or

python dcgan.py --mode generate --batch_size 128 --pretty

Result

Training progress:

Animation shows generated images during the training process of DCGAN over 100 epochs.