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

• Support parameters in cli *
• Add requirements.txt *
• Add Dockerfile if possible
• Multiple GPUs *
• Clean up code, remove unused code *

* not ready for conditional gan yet

• Example:

Fresh training

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --train_dir /path/to/train --validation_dir /path/to/validation/ --output_path /path/to/output/ --dim 64 --saving_step 300 --num_workers 8

Continued training:

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --train_dir /path/to/train --validation_dir /path/to/validation/ --output_path /path/to/output/ --dim 64 --saving_step 300 --num_workers 8 --restore_mode --start_iter 5000

• 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

Sample 1	Sample 2

• dining_room: 1
• bridge: 2
• restaurant: 3
• tower: 4

Sample 1	Sample 2

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

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

• igul222/improved_wgan_training
• caogang/wgan-gp
• LayerNorm