Process evaluateRejectionSampling in batches

[alexmill]

This commit modifies evaluateRejectionSampling so that, rather than iterating one-by-one through the dataset, rejections and rewards are evaluated in batches with numpy.

I haven't fully tested this with partial_fit methods or for situations when nchoices>2, but it works great for the policies I've tried. It's reduced the runtime for me by 100-fold. The benefits of batch processing obviously scale with the size of the batch, but improvements are noticeable even for small batches.

I might suggest adding tqdm as a dependency and adding a progress bar to big loops like this. In this case, one could replace

for i in range(X.shape[0]//batch_size+1):

with

for i in tqdm(range(X.shape[0]//batch_size+1), desc="Batches"):

This works great on my local version.

No pressure to merge, but this took me 10 minutes to write so I thought I'd make a pull request and share with the world (while I wait for my batches to run...). Cheers.

[david-cortes]

Thanks, merged. I'll change the docs too to reflect what it does. That progress bar looks good, but is there some way of making it an optional dependency?

[alexmill]

@david-cortes What is the minimum Python version required for this package? There is a way to only import a package if already installed, but the method for doing so depend on Python version.

[david-cortes]

There's no minimum required version, but I try to keep it as compatible as possible.