Note

I wrote this code before the official implementation got released. Now that the official implementation is part of Tensorflow, this codebase is not maintained anymore. Please refer to the official repo.

https://github.com/tensorflow/models/tree/master/research/real_nvp

real-nvp

Implementation of Real-NVP (https://arxiv.org/abs/1605.08803) in Tensorflow.

Started with code from PixelCNN++ by OpenAI (https://github.com/openai/pixel-cnn)

Sample usage:

1. Install Python3.
2. Create directories for downloading dataset and saving checkpoints.
3. Run train.py. '--nr_gpu', which denotes the number of GPUs to use, should be specified.

Sample usage:

$ CUDA_VISIBLE_DEVICES=1,2 python3 train.py --nr_gpu=2 --data_dir=download --save_dir=checkpoints --load_params=0 --save_interval=2

Sample image from the model trained on CIFAR10. The test NLL was 3.51.