Real NVP in PyTorch

Implementation of Real NVP in PyTorch. Based on the paper:

Density estimation using Real NVP
Laurent Dinh, Jascha Sohl-Dickstein, Samy Bengio
arXiv:1605.08803

Training script and hyperparameters designed to match the CIFAR-10 experiments described in Section 4.1 of the paper.

Usage

Environment Setup

1. Make sure you have Anaconda or Miniconda installed.
2. Clone repo with git clone https://github.com/chrischute/real-nvp.git rnvp.
3. Go into the cloned repo: cd rnvp.
4. Create the environment: conda env create -f environment.yml.
5. Activate the environment: source activate rnvp.

Train

1. Make sure you've created and activated the conda environment as described above.
2. Run python train.py -h to see options.
3. Run python train.py [FLAGS] to train. E.g., run python train.py for the default configuration, or run python train.py --gpu_ids=[0,1] --batch_size=128 to run on 2 GPUs instead of the default of 1 GPU.
4. At the end of each epoch, samples from the model will be saved to samples/epoch_N.png, where N is the epoch number.

One epoch takes about 4 minutes when using the default arguments and running on an NVIDIA Titan Xp card.

Samples

Epoch 5

Epoch 10

Epoch 15

Epoch 20

Epoch 25

Results

Bits per Dimension

Epoch	Train	Valid
5	3.97	3.98
10	3.76	3.76
15	3.69	3.74
20	3.65	3.70
25	3.62	3.74