need better sample data/tests #24
Description
(note: ignoring recommended feature request format for convenience)
copying from @mshvartsman's recent(ish) email:
I just pushed an initial stab at sparse TFA using a sort of hybrid approach between global-local shrinkage priors (specifically, horseshoe) and ARD/sparse Bayesian learning. Code: https://github.com/ContextLab/htfa/blob/sparse-tfa/htfa/sparse_tfa.py, minimal example https://github.com/ContextLab/htfa/blob/sparse-tfa/examples/Sparse%20TFA.ipynb.
The idea is that with something like this you don’t need to cross-validate over the number of factors, and I think it will be even more important when we go to more complex blobs. I’ve verified that on toy examples we get some sparsity and it can memorize (small/simple enough) data, but otherwise it’s not super stable / well-behaved and I’m not sure whether it’s because there’s a mistake in the code, or this setup is just poorly behaved numerically with Type-II MLE. Another pair of eyes on this would be awesome if you get around to it.
Also, I am starting to think that my super simple test examples aren’t very good as we get to higher complexity / capacity models — do you have any better ground truth test functions or synth examples that are useful for benchmarking and sanity checking things?
To do:
- Add/incorporate code for generating synthetic data for debugging (make sure we can recover known params)