AlainConnor / gan_practice

dcgan,wgan and improved wgan implementation by tensorflow

gan_practice

simple dcgan,wgan and improved wgan implementation by tensorflow, this is for practice purpose. tensorflow version=1.1.0

data preparation

download MNIST data from http://yann.lecun.com/exdb/mnist/ and unzip data to some dir,for example

data
|__ 10k-images-idx3-ubyte
|__ train-images-idx3-ubyte

training

python dcgan.py or wgan.py or wgan_gp.py \
                --data_path=data \
                --mode=train \
                --model_dir=models \
                --batch_size=100 \
                --epoch_num=100 \
                --rand_dim=128

generating

python dcgan.py or wgan.py or wgan_gp.py \
                --data_path=data \
                --mode=infer \
                --model_dir=models \
                --batch_size=100 \
                --epoch_num=100 \
                --rand_dim=128

some results

• dcgan 100 epoch results

* wgan 100 epoch results

* improved wgan 100 epoch results

reference

• Generative Adversarial Nets
• Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
• Wasserstein GAN
• Improved Training of Wasserstein GANs
• https://github.com/carpedm20/DCGAN-tensorflow
• https://github.com/shekkizh/WassersteinGAN.tensorflow
• https://github.com/igul222/improved_wgan_training