NVIDIA / NVIDIA/apex

Why using the gradient norm to control the dynamic loss scale

Open
#352 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
9k
Forks
1.5k
Avg merge
2d 4h
Merged PRs (30d)
3

Description

Hi everyone. Right now I am modifying the original FusedAdam( ) optimizer. I noticed that the dynamic scale factor is controlled by the norm of the gradients. Is there any specific reason why we are using this to control the scale factor, instead of using an element-wise metric? I mean can we check each element in the gradient and decay the scale factor whenever we find one Inf or NaN?

I would really appreciate it if you could help me with this! Thanks!

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Read the existing FusedAdam implementation to understand how the gradient norm controls the dynamic loss scale. Compare that behavior with element-wise Inf/NaN detection; the issue is done only once the project provides a maintainer-confirmed rationale or a clearly scoped change request.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.