Skip to content

[WIP] FIX Projected Gradient NMF stopping condition#2557

Closed
vene wants to merge 1 commit intoscikit-learn:masterfrom
vene:pgnmffix
Closed

[WIP] FIX Projected Gradient NMF stopping condition#2557
vene wants to merge 1 commit intoscikit-learn:masterfrom
vene:pgnmffix

Conversation

@vene
Copy link
Copy Markdown
Member

@vene vene commented Oct 29, 2013

I caught a bug in the PG NMF stopping condition. All in all it seems that this solver is not at all bad, but I need to understand it properly.

@vene
Copy link
Copy Markdown
Member Author

vene commented Oct 29, 2013

Benchmarks against Lars's PR.
text
faces

@coveralls
Copy link
Copy Markdown

Coverage Status

Coverage remained the same when pulling dadc2fc on vene:pgnmffix into 30eb78d on scikit-learn:master.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I find the code a bit hard to follow. I would rather define variables proj_grad_W and proj_grad_H and compute the Frobenius norms of those.

@vene
Copy link
Copy Markdown
Member Author

vene commented Nov 28, 2013

The non-negative least squares benchmark notebook version of the solver diverged a little bit. Most importantly the regularization is more properly done (but this makes the objective function and the parametrization different). Should I bring it in to this PR?

@ogrisel
Copy link
Copy Markdown
Member

ogrisel commented Nov 29, 2013

Could you summarize how it will impact the parametrization? Do you think there is a way to maintain backward compat with a deprecation cycle?

@vene
Copy link
Copy Markdown
Member Author

vene commented Jan 15, 2014

The default result of the estimator would change anyway in many cases because of the change in stopping conditions.

Should I just put the refactored solver in this PR to discuss it? Make a separate PR?

@ogrisel
Copy link
Copy Markdown
Member

ogrisel commented Jan 16, 2014

I am fine with bringing it into this PR to make it easier to see the changes.

@amueller
Copy link
Copy Markdown
Member

Should this be closed in light of #4852?

@agramfort
Copy link
Copy Markdown
Member

yes. Closing

@agramfort agramfort closed this Sep 22, 2015
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

8 participants