dimkal89 / glow Goto Github PK

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.

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.

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.

More samples can be found in the samples folder.