在Python中实现经典Runge-Kutta方案的方法有两种:here。第一个使用lambda函数,第二个没有它们。
哪一个会更快,为什么会更快?
答案 0 :(得分:4)
我调整了给定链接中的代码,并使用cProfile来比较两种技术:
import numpy as np
import cProfile as cP
def theory(t):
return (t**2 + 4.)**2 / 16.
def f(x, y):
return x * np.sqrt(y)
def RK4(f):
return lambda t, y, dt: (
lambda dy1: (
lambda dy2: (
lambda dy3: (
lambda dy4: (dy1 + 2*dy2 + 2*dy3 + dy4)/6
)( dt * f( t + dt , y + dy3 ) )
)( dt * f( t + dt/2, y + dy2/2 ) )
)( dt * f( t + dt/2, y + dy1/2 ) )
)( dt * f( t , y ) )
def test_RK4(dy=f, x0=0., y0=1., x1=10, n=10):
vx = np.empty(n+1)
vy = np.empty(n+1)
dy = RK4(f=dy)
dx = (x1 - x0) / float(n)
vx[0] = x = x0
vy[0] = y = y0
i = 1
while i <= n:
vx[i], vy[i] = x + dx, y + dy(x, y, dx)
x, y = vx[i], vy[i]
i += 1
return vx, vy
def rk4_step(dy, x, y, dx):
k1 = dx * dy(x, y)
k2 = dx * dy(x + 0.5 * dx, y + 0.5 * k1)
k3 = dx * dy(x + 0.5 * dx, y + 0.5 * k2)
k4 = dx * dy(x + dx, y + k3)
return x + dx, y + (k1 + k2 + k2 + k3 + k3 + k4) / 6.
def test_rk4(dy=f, x0=0., y0=1., x1=10, n=10):
vx = np.empty(n+1)
vy = np.empty(n+1)
dx = (x1 - x0) / float(n)
vx[0] = x = x0
vy[0] = y = y0
i = 1
while i <= n:
vx[i], vy[i] = rk4_step(dy=dy, x=x, y=y, dx=dx)
x, y = vx[i], vy[i]
i += 1
return vx, vy
cP.run("test_RK4(n=10000)")
cP.run("test_rk4(n=10000)")
得到了:
90006 function calls in 0.095 seconds
Ordered by: standard name
ncalls tottime percall cumtime percall filename:lineno(function)
1 0.000 0.000 0.095 0.095 <string>:1(<module>)
40000 0.036 0.000 0.036 0.000 untitled1.py:13(f)
1 0.000 0.000 0.000 0.000 untitled1.py:16(RK4)
10000 0.008 0.000 0.086 0.000 untitled1.py:17(<lambda>)
10000 0.012 0.000 0.069 0.000 untitled1.py:18(<lambda>)
10000 0.012 0.000 0.048 0.000 untitled1.py:19(<lambda>)
10000 0.009 0.000 0.027 0.000 untitled1.py:20(<lambda>)
10000 0.009 0.000 0.009 0.000 untitled1.py:21(<lambda>)
1 0.009 0.009 0.095 0.095 untitled1.py:28(test_RK4)
1 0.000 0.000 0.000 0.000 {method 'disable' of '_lsprof.Profiler' objects}
2 0.000 0.000 0.000 0.000 {numpy.core.multiarray.empty}
50005 function calls in 0.064 seconds
Ordered by: standard name
ncalls tottime percall cumtime percall filename:lineno(function)
1 0.000 0.000 0.064 0.064 <string>:1(<module>)
40000 0.032 0.000 0.032 0.000 untitled1.py:13(f)
10000 0.026 0.000 0.058 0.000 untitled1.py:43(rk4_step)
1 0.006 0.006 0.064 0.064 untitled1.py:51(test_rk4)
1 0.000 0.000 0.000 0.000 {method 'disable' of '_lsprof.Profiler' objects}
2 0.000 0.000 0.000 0.000 {numpy.core.multiarray.empty}
所以我会在&#34; lambda
&#34;中说function call overhead。实现使它变慢。
然而要注意,不管怎么说,我似乎已经失去了一些精确度,因为尽管彼此达成一致,结果仍然比示例中的结果更为明显:
>>> vx, vy = test_rk4()
>>> vy
array([ 1. , 1.56110667, 3.99324757, ..., 288.78174798,
451.27952013, 675.64427775])
>>> vx, vy = test_RK4()
>>> vy
array([ 1. , 1.56110667, 3.99324757, ..., 288.78174798,
451.27952013, 675.64427775])
答案 1 :(得分:3)
如果您使用Coconut转换程序预处理代码,它实现了尾部调用优化,那么它们是完全等效的(与未处理的更快版本一样快),因此您可以使用任何样式对你来说更方便。
# Save berna1111's code as rk4.coco; no modifications necessary.
$ coconut --target 3 rk4.coco & python3 rk4.py
50007 function calls in 0.055 seconds
Ordered by: standard name
ncalls tottime percall cumtime percall filename:lineno(function)
1 0.000 0.000 0.097 0.097 <string>:1(<module>)
40000 0.038 0.000 0.038 0.000 rk4.py:243(f)
1 0.000 0.000 0.000 0.000 rk4.py:246(RK4)
10000 0.007 0.000 0.088 0.000 rk4.py:247(<lambda>)
1 0.010 0.010 0.097 0.097 rk4.py:250(test_RK4)
1 0.000 0.000 0.097 0.097 {built-in method builtins.exec}
2 0.000 0.000 0.000 0.000 {built-in method numpy.core.multiarray.empty}
1 0.000 0.000 0.000 0.000 {method 'disable' of '_lsprof.Profiler' objects}
50006 function calls in 0.057 seconds
Ordered by: standard name
ncalls tottime percall cumtime percall filename:lineno(function)
1 0.000 0.000 0.057 0.057 <string>:1(<module>)
40000 0.030 0.000 0.030 0.000 rk4.py:243(f)
10000 0.019 0.000 0.049 0.000 rk4.py:265(rk4_step)
1 0.007 0.007 0.057 0.057 rk4.py:273(test_rk4)
1 0.000 0.000 0.057 0.057 {built-in method builtins.exec}
2 0.000 0.000 0.000 0.000 {built-in method numpy.core.multiarray.empty}
1 0.000 0.000 0.000 0.000 {method 'disable' of '_lsprof.Profiler' objects}
答案 2 :(得分:2)
@ berna1111和@ matt2000都是正确的。由于函数调用,lambda版本会产生额外的开销。尾调用优化将尾调用转换为while循环(即自动将lambda版本转换为while版本),从而消除函数调用开销。
请参阅https://stackoverflow.com/a/13592002/7421639了解为什么Python不会自动执行此优化,您必须使用像Coconut这样的工具来执行预处理过程。