Improved Training of Wasserstein GANs in Pytorch

This is a Pytorch implementation of gan_64x64.py from Improved Training of Wasserstein GANs.

Prerequisites

• Python >= 3.6

• Pytorch v0.4.0

• Numpy

• SciPy

• tensorboardX (installation here). It is very convenient to see costs and results during training with TensorboardX for Pytorch

• TensorFlow for tensorboardX

Model

• gan_train.py: This model is mainly based on GoodGenerator and GoodDiscriminator of gan_64x64.py model from Improved Training of Wasserstein GANs. It has been trained on LSUN dataset for around 100k iters.

• congan_train.py: ACGAN implementation, trained on 4 classes of LSUN dataset

Result

1. WGAN: trained on bedroom dataset (100k iters)

Sample 1 | Sample 2 :-------------------------:|:-------------------------:  | 

2. ACGAN: trained on 4 classes (100k iters)

• dining_room: 1

• bridge: 2

• restaurant: 3

• tower: 4

Sample 1 | Sample 2 :-------------------------:|:-------------------------:  | 

Testing

During the implementation of this model, we built a test module to compare the result between original model (Tensorflow) and our model (Pytorch) for every layer we implemented. It is available at compare-tensorflow-pytorch

TensorboardX

Results such as costs, generated images (every 200 iters) for tensorboard will be written to ./runs folder.

To display the results to tensorboard, run: tensorboard --logdir runs

Acknowledgements

• igul222/improved_wgan_training

• caogang/wgan-gp

• LayerNorm