Skip to content
Code for reproducing Flow ++ experiments
Branch: master
Clone or download
aravindsrinivas Merge pull request #2 from evanlohn/float64_prec
increased precision of floating point operations for numerical stability
Latest commit 737fadb Apr 27, 2019
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
flows initial import Jan 30, 2019
flows_celeba initial import Jan 30, 2019
flows_imagenet increased precision of floating point operations for numerical stability Apr 27, 2019
LICENSE.md Create LICENSE.md Jan 30, 2019
README.md Update README.md Feb 6, 2019

README.md

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

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).

You can’t perform that action at this time.