Glow in PyTorch

Implementation of Glow in PyTorch. Based on the paper:

Glow: Generative Flow with Invertible 1x1 Convolutions
Diederik P. Kingma, Prafulla Dhariwal
arXiv:1807.03039

Training script and hyperparameters designed to match the CIFAR-10 experiments described in Table 4 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/glow.git glow.
3. Go into the cloned repo: cd glow.
4. Create the environment: conda env create -f environment.yml.
5. Activate the environment: source activate glow.

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 to run on 2 GPUs instead of the default of 1 GPU. This will also double the batch size.
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.

A single epoch takes about 30 minutes with the default hyperparameters (K=32, L=3, C=512) on two 1080 Ti's.

Samples (K=16, L=3, C=512)

Epoch 10

Epoch 20

Epoch 30

Epoch 40

Epoch 50

Epoch 60

Epoch 70

Epoch 80

More samples can be found in the samples folder.

Results (K=32, L=3 C=512)

Bits per Dimension

Epoch	Train	Valid
10	3.64	3.63
20	3.51	3.56
30	3.46	3.53
40	3.43	3.51
50	3.42	3.50
60	3.40	3.51
70	3.39	3.49
80	3.38	3.49

Gradient Checkpointing

As pointed out by AlexanderMath, you can use gradient checkpointing to reduce memory consumption in the coupling layers. If interested, see this issue.