Exemple de profiling#
Profiling et fonction pdf. Le profiling est utilisé pour mesurer le temps que passe un programme dans chaque fonction.
Bizarrerie#
C’est un exemple qui m’a été envoyé par un étudiant pendant l’été pour montrer que la fonction pdf est plus lente qu’une réimplémentation simple qui fait à la même chose.
[1]:
import time
from scipy.stats import norm
import numpy as np
[2]:
debut = time.time()
for i in range(10**3):
norm(2, 3).pdf(4)
fin = time.time()
fin - debut
[2]:
0.9644453525543213
[3]:
def density(x, mean, sigma2):
return np.exp(-((x - mean) ** 2) / (2 * sigma2)) / (2 * np.pi * sigma2) ** 0.5
debut = time.time()
for i in range(10**3):
density(4, 2, 3)
fin = time.time()
fin - debut
[3]:
0.001481771469116211
Que se passe-t-il ?
Tout d’abord la fonction pdf comme toute les fonctions des librairies numériques sont optimisées pour le calcul sur des matrices ou des vecteurs et non sur des nombres. Pour la suite, on utilise un profileur.
Profiler#
[4]:
import cProfile, io, pstats, os, sys
def doprofile(func, filename, *l):
pr = cProfile.Profile()
pr.enable() # début du profiling
func(*l) # appel de la fonction
pr.disable() # fin du profiling
s = io.StringIO()
ps = pstats.Stats(pr, stream=s).sort_stats("cumulative")
ps.print_stats()
rem = os.path.normpath(os.path.join(os.getcwd(), "..", "..", ".."))
res = s.getvalue().replace(rem, "")
res = res.replace(sys.base_prefix, "").replace("\\", "/")
ps.dump_stats(filename)
return res
[5]:
import numpy
x = numpy.ones((10000000, 1)) * 4
x.shape
[5]:
(10000000, 1)
[6]:
debut = time.time()
y = norm.pdf(x)
fin = time.time()
print(fin - debut, y.shape, y[0])
0.6027283668518066 (10000000, 1) [0.00013383]
[17]:
import os
import scipy
path = os.path.normpath(os.path.join(scipy.__file__, "..", "..", ".."))
r = doprofile(norm.pdf, "pdf.prof", x)
print(r.replace(path, ""))
113 function calls in 0.450 seconds
Ordered by: cumulative time
ncalls tottime percall cumtime percall filename:lineno(function)
1 0.178 0.178 0.450 0.450 /site-packages/scipy/stats/_distn_infrastructure.py:1958(pdf)
1 0.000 0.000 0.142 0.142 /site-packages/scipy/stats/_continuous_distns.py:361(_pdf)
1 0.142 0.142 0.142 0.142 /site-packages/scipy/stats/_continuous_distns.py:300(_norm_pdf)
7 0.022 0.003 0.093 0.013 {built-in method numpy.core._multiarray_umath.implement_array_function}
1 0.000 0.000 0.048 0.048 <__array_function__ internals>:177(place)
1 0.000 0.000 0.048 0.048 /site-packages/numpy/lib/function_base.py:1912(place)
1 0.048 0.048 0.048 0.048 {built-in method numpy.core._multiarray_umath._insert}
1 0.028 0.028 0.028 0.028 /site-packages/scipy/stats/_distn_infrastructure.py:975(_support_mask)
1 0.000 0.000 0.022 0.022 <__array_function__ internals>:177(putmask)
2 0.000 0.000 0.022 0.011 /site-packages/numpy/core/fromnumeric.py:69(_wrapreduction)
2 0.021 0.011 0.021 0.011 {method 'reduce' of 'numpy.ufunc' objects}
1 0.000 0.000 0.021 0.021 /site-packages/scipy/stats/_distn_infrastructure.py:559(argsreduce)
1 0.000 0.000 0.011 0.011 <__array_function__ internals>:177(all)
1 0.000 0.000 0.011 0.011 /site-packages/numpy/core/fromnumeric.py:2406(all)
1 0.000 0.000 0.011 0.011 <__array_function__ internals>:177(any)
1 0.000 0.000 0.011 0.011 /site-packages/numpy/core/fromnumeric.py:2307(any)
1 0.000 0.000 0.010 0.010 /site-packages/scipy/stats/_distn_infrastructure.py:604(<listcomp>)
2 0.010 0.005 0.010 0.005 {method 'ravel' of 'numpy.ndarray' objects}
1 0.000 0.000 0.000 0.000 <__array_function__ internals>:177(broadcast_arrays)
1 0.000 0.000 0.000 0.000 /site-packages/numpy/lib/stride_tricks.py:480(broadcast_arrays)
1 0.000 0.000 0.000 0.000 /site-packages/numpy/lib/stride_tricks.py:546(<listcomp>)
3 0.000 0.000 0.000 0.000 /site-packages/numpy/lib/stride_tricks.py:340(_broadcast_to)
2 0.000 0.000 0.000 0.000 /site-packages/numpy/core/_ufunc_config.py:32(seterr)
1 0.000 0.000 0.000 0.000 <__array_function__ internals>:177(atleast_1d)
1 0.000 0.000 0.000 0.000 /site-packages/numpy/core/_ufunc_config.py:434(__exit__)
1 0.000 0.000 0.000 0.000 /site-packages/numpy/core/_ufunc_config.py:429(__enter__)
1 0.000 0.000 0.000 0.000 /site-packages/numpy/core/shape_base.py:23(atleast_1d)
1 0.000 0.000 0.000 0.000 <__array_function__ internals>:177(shape)
1 0.000 0.000 0.000 0.000 {built-in method numpy.zeros}
1 0.000 0.000 0.000 0.000 /site-packages/numpy/lib/stride_tricks.py:416(_broadcast_shape)
2 0.000 0.000 0.000 0.000 /site-packages/numpy/core/_ufunc_config.py:131(geterr)
3 0.000 0.000 0.000 0.000 {built-in method builtins.any}
1 0.000 0.000 0.000 0.000 /site-packages/numpy/lib/stride_tricks.py:538(<listcomp>)
2 0.000 0.000 0.000 0.000 {built-in method numpy.seterrobj}
3 0.000 0.000 0.000 0.000 /site-packages/numpy/lib/function_base.py:346(iterable)
4 0.000 0.000 0.000 0.000 {built-in method numpy.geterrobj}
9 0.000 0.000 0.000 0.000 /site-packages/numpy/lib/stride_tricks.py:345(<genexpr>)
1 0.000 0.000 0.000 0.000 {built-in method numpy.asarray}
1 0.000 0.000 0.000 0.000 {built-in method builtins.all}
6 0.000 0.000 0.000 0.000 {built-in method numpy.array}
2 0.000 0.000 0.000 0.000 /site-packages/numpy/core/fromnumeric.py:70(<dictcomp>)
1 0.000 0.000 0.000 0.000 {method 'reshape' of 'numpy.ndarray' objects}
1 0.000 0.000 0.000 0.000 /site-packages/numpy/core/fromnumeric.py:1965(shape)
3 0.000 0.000 0.000 0.000 /site-packages/numpy/lib/stride_tricks.py:542(<genexpr>)
1 0.000 0.000 0.000 0.000 {built-in method numpy.promote_types}
1 0.000 0.000 0.000 0.000 /site-packages/numpy/core/_ufunc_config.py:425(__init__)
1 0.000 0.000 0.000 0.000 /site-packages/scipy/stats/_distn_infrastructure.py:941(_argcheck)
3 0.000 0.000 0.000 0.000 /site-packages/numpy/lib/stride_tricks.py:25(_maybe_view_as_subclass)
2 0.000 0.000 0.000 0.000 {method 'items' of 'dict' objects}
1 0.000 0.000 0.000 0.000 /site-packages/numpy/core/fromnumeric.py:2302(_any_dispatcher)
1 0.000 0.000 0.000 0.000 /site-packages/numpy/core/multiarray.py:1106(putmask)
1 0.000 0.000 0.000 0.000 /site-packages/numpy/core/fromnumeric.py:1961(_shape_dispatcher)
2 0.000 0.000 0.000 0.000 {built-in method builtins.len}
3 0.000 0.000 0.000 0.000 {method '__exit__' of 'numpy.nditer' objects}
3 0.000 0.000 0.000 0.000 {built-in method builtins.iter}
1 0.000 0.000 0.000 0.000 /site-packages/numpy/lib/function_base.py:1908(_place_dispatcher)
2 0.000 0.000 0.000 0.000 {method 'append' of 'list' objects}
1 0.000 0.000 0.000 0.000 /site-packages/scipy/stats/_distn_infrastructure.py:953(_get_support)
2 0.000 0.000 0.000 0.000 {built-in method builtins.isinstance}
1 0.000 0.000 0.000 0.000 {method 'disable' of '_lsprof.Profiler' objects}
2 0.000 0.000 0.000 0.000 {built-in method numpy.asanyarray}
1 0.000 0.000 0.000 0.000 <string>:2(_parse_args)
1 0.000 0.000 0.000 0.000 /site-packages/numpy/core/shape_base.py:19(_atleast_1d_dispatcher)
1 0.000 0.000 0.000 0.000 /site-packages/numpy/core/fromnumeric.py:2401(_all_dispatcher)
1 0.000 0.000 0.000 0.000 /site-packages/numpy/lib/stride_tricks.py:476(_broadcast_arrays_dispatcher)
[8]:
def density(x, mean, sigma2):
return np.exp(-((x - mean) ** 2) / (2 * sigma2)) / (2 * np.pi * sigma2) ** 0.5
debut = time.time()
y = density(x, 0.0, 1.0)
fin = time.time()
print(fin - debut, y.shape, y[0])
0.1882781982421875 (10000000, 1) [0.00013383]
[9]:
r = doprofile(density, "pdf.prof", x, 0, 1)
print(r)
2 function calls in 0.177 seconds
Ordered by: cumulative time
ncalls tottime percall cumtime percall filename:lineno(function)
1 0.177 0.177 0.177 0.177 /tmp/ipykernel_29119/200996087.py:1(density)
1 0.000 0.000 0.000 0.000 {method 'disable' of '_lsprof.Profiler' objects}
Quand on regarde le code de la fonction, on s’aperçoit que la fonction perd du temps dans argsreduce. Elle fait aussi d’autres choses comme regarder les valeurs manquantes. En guise de conclusion, lorsqu’une fonction gère trop de cas particuliers (type, valeurs), elle est nécessairement plus lente qu’une fonction qu’on implémente soi-même.
[ ]:
[ ]: