Numba Kernel

[AxelGiottonini]

Hello!

First of all, thanks for this tool!

I noticed that using the current library was very slow on large datasets and therefore I implemented a version of the base Kernel using Numba here.

I compared the NbKernel against the base Kernel

%%timeit

kernel = NbKernel(
    kernel="rbf",
    sigma=3,
    degree=2,
    coef0=1,
    normalize=True
)

X = np.random.randn(10000, 20)
Y = np.random.randn(10000, 20)

kernel.matrix(X, Y)

8.74 s ± 371 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

%%time

kernel = Kernel(
    kernel="rbf",
    sigma=3,
    degree=2,
    coef0=1,
    normalize=True
)

X = np.random.randn(10000, 20)
Y = np.random.randn(10000, 20)

kernel.matrix(X, Y)

CPU times: user 6min 48s, sys: 1.46 s, total: 6min 50s
Wall time: 6min 52s

Are you interested in maintaining the library and optimizing it? If that's the case and my implementaiton fits your idea, I'm happy to cooperate!