tensorflow / tensorflow/probability

bfgs fails with batched initial inverse Hessians

Open
#667 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

When using the bfgs optimizer in batch mode, there is a bug when providing batched initial inverse hessians. This is useful to continue an incomplete batched optimization.

I tracked the bug to the _inv_hessian_control_inputs function in bfgs.py. The two calls to tf.Assert have boolean tensors instead of scalars as their first arguments. tf.Assert can only take a single boolean scalar.

This can be easily fixed with tf.reduce_all wrapping the tensors is_positive_definite and is_symmetric.

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 in bfgs.py at the _inv_hessian_control_inputs function and inspect the two tf.Assert calls for batched initial inverse Hessians. Verify that batched positive-definiteness and symmetry checks are accepted, then run the relevant BFGS test suite to confirm the optimizer can continue batched optimization without assertion errors.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, tensorflow
Domain
machine-learning
Issue type
Bug
Difficulty
2/5
Estimated time
1-3 hours
Activity status
Stale
Clarity
Clearly specified
Newbie friendliness
55/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.