Make the utils.isnan function inline

[alexhenrie]

Bayes net training is about 35% faster when isnan is inlined, which makes sense because npy_isnan was designed to be an inline function.

Test program:

import numpy as np
from neurtu import delayed, Benchmark
from pomegranate import BayesianNetwork
from random import random

X = np.round(np.random.rand(10000, 16))
train = delayed(BayesianNetwork)().from_samples(X)

print(Benchmark(wall_time=True, cpu_time=True, repeat=3)(train))

Non-inline isnan:

      wall_time  cpu_time                                                                                                     
mean   8.369287  8.361042
max    8.488417  8.471947
std    0.103549  0.112760

Inline isnan:

      wall_time  cpu_time                                                                                                     
mean   5.361331  5.360565
max    5.379823  5.376542
std    0.030915  0.019036

[jmschrei]

I can reproduce this. Nice find.

[alexhenrie]

Thank you!!!