Increase training and inference performance for GAK kernel

[lejafar]

Hi! I've been using the TimeSeriesSVC with the gak kernel for a while now and I've made some changes that increased the performance quite a bit, the most time consuming part in my case was computing the kernel and its there where I've made some changes

Changes that apply to training:

To compute the self-similarity, scikit-learn evaluates the callable kernel using a reference of X as the second argument, in the current implementation will check if the second argument is None which it isn't, hence it computes everything twice

Changes that apply to inference:

When doing prediction, scikit-learn evaluates the callable kernel for every point (X_fit) seen up until then, but the only thing libsvm really needs is the distance to its support vectors, hence we can omit computing the others.

[rtavenar]

Thanks for your PR, this is a good idea indeed.

[rtavenar]

Thanks for your PR, this is a good idea indeed.

Could you please answer my 2 questions so that I can make my opinion on these two technical points?

Thanks,
Romain

[rtavenar]

Woops sorry for double posting my review...

[lejafar]

@rtavenar anything else?

[rtavenar]

Hi @lejafar

Sorry it took me so long to get back to you. Your suggestion is still welcome but the codebase has changed quite a lot since you posted it. Specifically, we are working on a future release in the dev branch.

So, my question is : would you be willing to close this PR and open a new one based on the current dev branch ?

The place where your hack could be inserted is probably around:
https://github.com/rtavenar/tslearn/blob/dev/tslearn/svm.py#L58-L60