Speedup achieved by altering NMF parameters during initialization

[fdeguire03]

The end of the GreedyROI function in initialization.py uses sklearn.decomposition.NMF to estimate the background component in each patch:

model = NMF(n_components=nb, init='nndsvdar')
b_in = model.fit_transform(np.maximum(res, 0)).astype(np.float32)
f_in = model.components_.astype(np.float32)

I have found that the parameters init=random and max_iter=25 greatly improve the runtime (often by 2-3x) for this method without any detriment to quality (both methods give roughly the same reconstruction error in model.reconstruction_err_). Additionally, since initialize_components function immediately refines the results of GreedyROI using HALS, so using fewer iterations should not be an issue.

Would the maintainers consider switching to these NMF parameters by default or at least making them configurable via params.py?

[pgunn]

We've exposed more parameters in the past when people requested them; I don't see any reason we couldn't do that here.

[pgunn]

@kushalkolar I'm wondering if you have any thoughts on changing the defaults

[kushalkolar]

Sure expose the params but I wouldn't change the defaults since that can mess with other people's known params.

[pgunn]

@fdeguire03 I'll try to get these exposed by the next release; likely it'll be in dev within the week.