Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design

This repository contains Tensorflow implementation of experiments from the paper Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design - Jonathan Ho, Xi Chen, Aravind Srinivas, Yan Duan, Pieter Abbeel

Dependencies

• python3.6
• Tensorflow v1.10.1
• horovod v0.14.1

Horovod GPU setup instructions

Usage Instructions

We trained our models using 8 GPUs with data-parallelism using Horovod.

CIFAR 10

mpiexec -n 8 python3.6 run_cifar.py

Imagenet

Data for Imagenet Experiments:

Script to create dataset here

Imagenet 32x32

mpiexec -n 8 python3.6 -m flows_imagenet.launchers.imagenet32_official

Imagenet 64x64

mpiexec -n 6 python3.6 -m flows_imagenet.launchers.imagenet64_official
mpiexec -n 6 python3.6 -m flows_imagenet.launchers.imagenet64_5bit_official

CelebA-HQ 64x64

Data:

Download links in README

mpiexec -n 8 python3.6 -m flows_celeba.launchers.celeba64_5bit_official
mpiexec -n 8 python3.6 -m flows_celeba.launchers.celeba64_3bit_official

Contact

Please open an issue

Credits

flowpp was originally developed by Jonathan Ho (UC Berkeley), Peter Chen (UC Berkeley / covariant.ai), Aravind Srinivas (UC Berkeley), Yan Duan (covariant.ai), and Pieter Abbeel (UC Berkeley / covariant.ai).