pyro-ppl/pyro

FR: Volume Preserving Real NVP Variant (GIN)

Open

#2,747 opened on Jan 22, 2021

View on GitHub
 (4 comments) (0 reactions) (0 assignees)Python (981 forks)batch import
enhancementhelp wanted

Repository metrics

Stars
 (8,211 stars)
PR merge metrics
 (Avg merge 10d 19h) (1 merged PR in 30d)

Description

Hi all,

I noticed that Real-NVP has an option for "Volume Preserving", but this just omits the scale factors, and performs the NICE transform. Would it be possible to implement a more powerful volume preserving transformation (I need it for a specific purpose) by simply modifying Real-NVP so the final scale variable is the negative sum of the others?

The very simple idea can be found in this ICLR2020 paper: https://arxiv.org/abs/2001.04872

Similarly, it should be possible to do this for IAF.

It should be easy for someone who is familiar with the code already.

Contributor guide

Research direction
Implement the volume preserving transformation by modifying the Real NVP coupling layer so that the final scale variable equals the negative sum of the other scales, as described in the ICLR2020 paper (arXiv:2001.04872). Extend similarly to IAF.
Tech stack
python
Domain
machine learning
Issue type
Feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-2 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Clear
Prerequisites
PyTorchPyronormalizing flows
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
50