tensorflow / tensorflow/probability

Suggestion: do not reoptimize already converged points

Open
#1,707 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Jupyter Notebook
Stars
4.4k
Forks
1.1k
PR merge metrics
No merged PRs in 30d

Description

Similar to the idea from https://github.com/tensorflow/probability/pull/928, it seems like internally the (l)bgfs optimisers will go through points that already have reached convergence if any other points in the problem have not. The linked PR allows one to re-enter the optimiser with an initial state, allowing one to remove those already converged points manually.

I was wondering if doing that internally would be an option, that is skipping computations for batches that are already converged, or if that would be too complicated/slow for the general use case to do that internally and automatically.

Contributor guide

Open the contributing guide

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

Start by reviewing the linked TensorFlow Probability pull request 928 and the internal L-BFGS optimizers discussed in the issue. Determine whether already-converged batches can be skipped without affecting unfinished points, then document the behavior and add coverage demonstrating that converged points are not reoptimized.

Written by the indexing model from the issue text.

Assessment

Tech stack
tensorflow
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.