Nice figure! Let me see with the others if this should go in.

Could you run on a wikipedia text example?

Here is a benchmark on word counts for the first paragraph of 500k wikipedia articles. To speed up calculations, I used a HashingVectorizer with 4096 features. The KL divergence is calculated on 50k test samples.

Can you do a draft PR just adding the code to scikit-learn, so that people can have a quick look at how complicated the code would be.

Don't worry about test, documentation, or full API compatibility for now. Just add the object (which I think should be called MiniBatchNMF).

would be nice to see more full benchmarks but I think in general this would be useful.

vary n_features and n_samples?

@GaelVaroquaux , @amueller I have synchronized with master the code proposed by @pcerda in #13386. I was unable to access the data set used in the previous comment, I used the "Blog Authorship Corpus" then.
The attached plot benchmarks n_features, n_samples and n_components with the scikit-learn implemented version of Non-Negative Matrix Factorization (NMF) and the online one proposed by @pcerda.

The code used to produce the plot is a variation on the topic extraction example. It is available here.
As a reminder, the paper to be referred to is Online algorithms for nonnegative matrix factorization with the Itakura-Saito divergence, A Lefevre, F Bach, C Févotte, 2011.

Comments?
Thanks!

The speed up is impressive. Thank you for highlighting it.

@cmarmo : can you push a PR with the new code. So that we can review and iterate on this.

Thanks!!

@cmarmo : can you push a PR with the new code. So that we can review and iterate on this.

I'm on it: just trying to get rid of the NMFOriginal.py file. Patricio kept it in order to guarantee the comparison with the original implementation, I'm trying to reproduce the original behavior in the same code. I can open a WIP PR if you think it's worth it, but I would rather if you wait before reviewing the code. Thanks!

@cmarmo in the benchmarks you made above, you display the convergence time. What is the convergence criterion and do both versions converge to a solution of similar quality ?

@jeremiedbb I'm working to add this information to the plot. Thanks for your comment.

Hi @jeremiedbb and thanks for your patience, my plot behaved fuzzy probably because someone else was heavyly computing on the machine at the same time... waiting for the new version let me just say that, if I understand correctly, the convergence is computed here

@jeremiedbb this new plot and my previous comment, are they answering your question?

So in all your benchmarks, online NMF converges faster to a better solution. Pretty convincing !

As I'm starting to play with the source, for the sake of reproducibility this plot and this one have been produced with this version of the nmf code.