Deep Convolutional Generative Adversarial Networks (DCGAN)

This repository introduces the TensorFlow 2 implementation of DCGAN described in UNSUPERVISED REPRESENTATION LEARNING WITH DEEP CONVOLUTIONAL GENERATIVE ADVERSARIAL NETWORKS (Radford et al. 2015)

Requirements

• Python 3.7 or later
• Tensorflow 2.2.0
• numpy
• opencv
• matplotlib

CelebA Dataset

The model in this repository was trained on CelebA Dataset, which is a large-scale face dataset containing 202.599 celebrity face images. To reproduce the result, the Align&Cropped Images of the CelebA Dataset are required. Please refer to Large-scale CelebFaces Attributes (CelebA) Dataset for more information and downloading.

Training

To train DCGAN model, use the following command:

python run.py --dataset_dir path \ 
              --checkpoint_dir path --progress_dir path \
              --latent_dim int --test_size int \
              --batch_size int --lr float --epochs int

where:

• --checkpoint_dir: Path to save checkpoint. Default="./model/checkpoint"
• --progress_dir: Path to write training progress image. Default="./data/face_gan"
• --dataset_dir: Path to dataset.
• --latent_dim: Latent space dimension. Default=100
• --test_size: Number of test images during training is equal test_size^2. Default=4
• --batch_size: Number of training steps per epoch. Default=100
• --lr: Learning rate. Default=0.0002
• --epochs: Number of epochs for training. Default=20

Training progress

This GIF illustrates the generator improvement during the training

Results

Samples from latent space	Faces linear transformation