Automatically force_all_finite=False in non-initial warm_start runs

[jnothman]

For estimators that have a warm_start option, we generally expect that the same data is passed in for each call to fit as was done in the first. I think we can save a little on runtime by using check_array with force_all_finite=False in the case that warm_start=True and the model has already been fitted.

[cr458]

Can I take this?

[jnothman]

I'd be interested in seeing if other core devs think we should do this; there may not be consensus. But with that caveat you could certainly propose a PR. you could test it by sneaking a np.inf into the data before a second fit.

[mohamed-ali]

@jnothman, what's the status on this? I can work on it, if @cr458 is not available.

[jnothman]

we've still not heard if other core devs have an opinion. so you can offer a PR, but with no assurance it will ever get merged

[Geoffrey-Paris]

Hello @jnothman

I worked on this issue and i submitted a PR for this.
I know this issue is quite old and it needs approval from other devs but here's what I've done so far. I tried to make the most generic change by updating the _validate_data method of the BaseEstimator class but there are still some other calls to check_array outside the _validate_data method.

[amueller]

I think warm_start has inconsistent semantics across modules. In linear models, you can warm-start with different datasets and potentially get a speed benefit but essentially an equivalent solution, while for trees, you have to warm-start on the same dataset. So I don't think we can actually do this consistently and probably first maybe clarify our warm-start semantics?