experimental_feature.py

import chainer
from chainer import links as L
from chainer import utils
import warnings
import time
import numpy as np
import cupy as cp


warnings.simplefilter('ignore')


def evaluate(f):
    st = time.time()
    for _ in range(100):
        f()
    ed = time.time()
    print(ed - st)


def linear_function(xp, experimental=None):
    if experimental is not None:
        if experimental == 'no_api_name':
            utils.experimental()
        else:
            utils.experimental(experimental)

    x = chainer.Variable(xp.random.uniform(-1, 1, (128, 128)).astype(np.float32))
    l = L.Linear(128, 128)
    if xp is cp:
        l.to_gpu()
    l(x)


def f():
    pass

def g():
    utils.experimental('g')

def h():
    utils.experimental()


evaluate(f)
evaluate(g)
evaluate(h)
print()
evaluate(lambda : linear_function(np))
evaluate(lambda : linear_function(np, 'foo'))
evaluate(lambda : linear_function(np, 'no_api_name'))
print()
linear_function(cp)  # caching                                                                                                                                                                                                                                                                                               
evaluate(lambda : linear_function(cp))
evaluate(lambda : linear_function(cp, 'foo'))
evaluate(lambda : linear_function(cp, 'no_api_name'))

'''
6.67572021484375e-06
0.000202178955078125
0.05139327049255371

0.1634047031402588
0.1674811840057373
0.22064852714538574

0.11320042610168457
0.113677978515625
0.14753389358520508
'''