python - 它应该更快，cProfile说它更快，但程序实际上运行得更慢

Question

希望你能帮忙，因为这个让我很头疼。

我正在用 python(3.3) 开发一个小型捕食者-猎物模拟，它使用一个简单的前馈神经网络。今天，我将执行大脑每个“滴答声”的函数从纯 python 更改为 numpy 数组，以便在我使用更大的大脑时对其进行优化。

我用 cProfile 检查了(整个主程序“循环”的)性能速度，正如我所料，“tick”函数(在 brain 类中)更快。但是，在运行该程序时，我注意到实际上使用 numpy 的速度较慢(100 次循环在 ~12 秒内 vs 100 次循环在 ~9 秒内)。

为什么会这样？这是代码:

原始实现:

class Brain:

def __init__(self, inputs, hidden, outputs):
    self.input_num = inputs
    self.hidden_num = hidden
    self.output_num = outputs

    self.h_weight = []
    self.o_weight = []
    for _ in range(self.input_num * self.hidden_num):
        self.h_weight.append(random.random()*2-1)
    for _ in range(self.hidden_num * self.output_num):
        self.o_weight.append(random.random()*2-1)

def tick(self):
    input_num = self.input_num 
    hidden_num = self.hidden_num
    output_num = self.output_num

    hidden = [0]*hidden_num
    output = [0]*output_num

    inputs = self.input
    h_weight = self.h_weight
    o_weight = self.o_weight

    e = math.e

    count = -1
    for x in range(hidden_num):
        temp = 0
        for y in range(input_num):
            count += 1
            temp -= inputs[y] * h_weight[count]
        hidden[x] = 1/(1+e**(temp))  

    count = -1      
    for x in range(output_num):
        temp = 0 
        for y in range(hidden_num):
            count += 1 
            temp -= hidden[y] * o_weight[count]
        output[x] = 1/(1+e**(temp))  

    self.output = output

新实现(使用 numpy):

class Brain:

def __init__(self, inputs, hidden, outputs):
    self.input_num = inputs
    self.hidden_num = hidden
    self.output_num = outputs


    self.h_weights = random.random((self.hidden_num, self.input_num)) 
    self.o_weights = random.random((self.output_num, self.hidden_num))

    self.h_activation = zeros((self.hidden_num, 1), dtype=float)
    self.o_activation = zeros((self.output_num, 1), dtype=float)

    self.i_output = zeros((self.input_num, 1), dtype=float)      
    self.h_output = zeros((self.hidden_num, 1), dtype=float)  
    self.o_output = zeros((self.output_num, 1), dtype=float)

def tick(self):
    i_output = self.input
    h_weights = self.h_weights
    o_weights = self.o_weights

    h_activation = dot(h_weights, i_output)
    h_output = tanh(h_activation)

    o_activation = dot(o_weights, h_output)
    o_output = tanh(o_activation)
    self.output = o_output  

这里是程序的主循环，我已经计时(忽略其他函数，它们对“brain.tick()”函数没有影响)。它在另一个类中，这也是无关紧要的:

    def update(self):
    GUI.update()

    if not self.pause:   
        self.tick += 1

        if self.tick % 1000 == 0:
            if self.globalSelection:
                self.newGeneration(self.creatures)
            else:
                for specie in self.species:
                    if not specie.isPlant:
                        creatureList = [creature for creature in self.creatures if                creature.specie == specie] 
                        self.newGeneration(creatureList)   


        for creature in self.creatures:
            if not creature.specie.isPlant:
                if self.useHunger and creature.hunger < 1:
                    creature.hunger += 1/240    
                creature.setInputs()
                creature.brain.tick()
                creature.move(creature.brain.output[0]*50-25, creature.brain.output[1]*8)
                creature.interactions()

    threading.Timer(self.tickrate, self.update).start()

现在大脑设置为 5 个输入、200 个隐藏和 2 个输出，只是为了测试速度。以下是 cProfile 的结果:

原文:

         3312 function calls in 0.094 seconds



Ordered by: standard name

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.000    0.000    0.094    0.094 <string>:1(<module>)
      200    0.006    0.000    0.007    0.000 __init__.py:263(move)
      200    0.024    0.000    0.024    0.000 __init__.py:286(setInputs)
      200    0.001    0.000    0.001    0.000 __init__.py:360(interactions)
      200    0.033    0.000    0.033    0.000 __init__.py:415(tick)
        1    0.005    0.005    0.094    0.094 __init__.py:46(update)
        1    0.007    0.007    0.020    0.020 __init__.py:471(update)
        1    0.000    0.000    0.000    0.000 _weakrefset.py:79(add)
        3    0.000    0.000    0.000    0.000 threading.py:127(__init__)
        1    0.000    0.000    0.000    0.000 threading.py:160(_release_save)
        1    0.000    0.000    0.000    0.000 threading.py:163(_acquire_restore)
        1    0.000    0.000    0.000    0.000 threading.py:166(_is_owned)
        1    0.000    0.000    0.000    0.000 threading.py:175(wait)
        2    0.000    0.000    0.000    0.000 threading.py:297(__init__)
        1    0.000    0.000    0.000    0.000 threading.py:305(is_set)
        1    0.000    0.000    0.000    0.000 threading.py:325(wait)
        1    0.000    0.000    0.000    0.000 threading.py:507(_newname)
        1    0.000    0.000    0.000    0.000 threading.py:534(__init__)
        1    0.000    0.000    0.000    0.000 threading.py:577(start)
        1    0.000    0.000    0.000    0.000 threading.py:775(daemon)
        1    0.000    0.000    0.000    0.000 threading.py:810(__init__)
        1    0.000    0.000    0.000    0.000 threading.py:887(current_thread)
      225    0.002    0.000    0.002    0.000 {built-in method aacircle}
      200    0.001    0.000    0.001    0.000 {built-in method aaline}
      200    0.000    0.000    0.000    0.000 {built-in method abs}
        4    0.000    0.000    0.000    0.000 {built-in method allocate_lock}
      200    0.001    0.000    0.001    0.000 {built-in method atan2}
      400    0.001    0.000    0.001    0.000 {built-in method cos}
        1    0.000    0.000    0.094    0.094 {built-in method exec}
      225    0.001    0.000    0.001    0.000 {built-in method filled_circle}
        4    0.000    0.000    0.000    0.000 {built-in method filled_polygon}
        1    0.000    0.000    0.000    0.000 {built-in method get_ident}
        1    0.000    0.000    0.000    0.000 {built-in method get_pressed}
        1    0.000    0.000    0.000    0.000 {built-in method get}
        2    0.000    0.000    0.000    0.000 {built-in method len}
      200    0.000    0.000    0.000    0.000 {built-in method radians}
        1    0.000    0.000    0.000    0.000 {built-in method round}
      400    0.001    0.000    0.001    0.000 {built-in method sin}
        1    0.000    0.000    0.000    0.000 {built-in method start_new_thread}
        1    0.006    0.006    0.006    0.006 {built-in method update}
        5    0.000    0.000    0.000    0.000 {method 'acquire' of '_thread.lock' objects}
        1    0.000    0.000    0.000    0.000 {method 'add' of 'set' objects}
        1    0.000    0.000    0.000    0.000 {method 'append' of 'list' objects}
        2    0.000    0.000    0.000    0.000 {method 'blit' of 'pygame.Surface' objects}
        1    0.000    0.000    0.000    0.000 {method 'disable' of '_lsprof.Profiler' objects}
        1    0.001    0.001    0.001    0.001 {method 'fill' of 'pygame.Surface' objects}
        8    0.000    0.000    0.000    0.000 {method 'random_sample' of 'mtrand.RandomState' objects}
        2    0.000    0.000    0.000    0.000 {method 'release' of '_thread.lock' objects}
        3    0.000    0.000    0.000    0.000 {method 'render' of 'pygame.font.Font' objects}

新:

                3322 function calls in 0.068 seconds



 Ordered by: standard name

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.000    0.000    0.068    0.068 <string>:1(<module>)
      200    0.005    0.000    0.006    0.000 __init__.py:265(move)
      200    0.022    0.000    0.023    0.000 __init__.py:288(setInputs)
      200    0.001    0.000    0.001    0.000 __init__.py:362(interactions)
      200    0.005    0.000    0.014    0.000 __init__.py:417(tick)
        1    0.005    0.005    0.068    0.068 __init__.py:47(update)
        1    0.005    0.005    0.019    0.019 __init__.py:473(update)
        1    0.000    0.000    0.000    0.000 _weakrefset.py:79(add)
        3    0.000    0.000    0.000    0.000 threading.py:127(__init__)
        1    0.000    0.000    0.000    0.000 threading.py:160(_release_save)
        1    0.000    0.000    0.000    0.000 threading.py:163(_acquire_restore)
        1    0.000    0.000    0.000    0.000 threading.py:166(_is_owned)
        1    0.000    0.000    0.000    0.000 threading.py:175(wait)
        2    0.000    0.000    0.000    0.000 threading.py:297(__init__)
        1    0.000    0.000    0.000    0.000 threading.py:305(is_set)
        1    0.000    0.000    0.000    0.000 threading.py:325(wait)
        1    0.000    0.000    0.000    0.000 threading.py:507(_newname)
        1    0.000    0.000    0.000    0.000 threading.py:534(__init__)
        1    0.000    0.000    0.000    0.000 threading.py:577(start)
        1    0.000    0.000    0.000    0.000 threading.py:775(daemon)
        1    0.000    0.000    0.000    0.000 threading.py:810(__init__)
        1    0.000    0.000    0.000    0.000 threading.py:887(current_thread)
      225    0.002    0.000    0.002    0.000 {built-in method aacircle}
      200    0.001    0.000    0.001    0.000 {built-in method aaline}
      200    0.001    0.000    0.001    0.000 {built-in method abs}
        4    0.000    0.000    0.000    0.000 {built-in method allocate_lock}
      200    0.000    0.000    0.000    0.000 {built-in method atan2}
      400    0.001    0.000    0.001    0.000 {built-in method cos}
      400    0.008    0.000    0.008    0.000 {built-in method dot}
        1    0.000    0.000    0.068    0.068 {built-in method exec}
      225    0.001    0.000    0.001    0.000 {built-in method filled_circle}
        4    0.000    0.000    0.000    0.000 {built-in method filled_polygon}
        1    0.000    0.000    0.000    0.000 {built-in method get_ident}
        1    0.000    0.000    0.000    0.000 {built-in method get_pressed}
        1    0.000    0.000    0.000    0.000 {built-in method get}
        2    0.000    0.000    0.000    0.000 {built-in method len}
      200    0.000    0.000    0.000    0.000 {built-in method radians}
        1    0.000    0.000    0.000    0.000 {built-in method round}
      400    0.001    0.000    0.001    0.000 {built-in method sin}
        1    0.000    0.000    0.000    0.000 {built-in method start_new_thread}
        1    0.005    0.005    0.005    0.005 {built-in method update}
        5    0.000    0.000    0.000    0.000 {method 'acquire' of '_thread.lock' objects}
        1    0.000    0.000    0.000    0.000 {method 'add' of 'set' objects}
        1    0.000    0.000    0.000    0.000 {method 'append' of 'list' objects}
        2    0.000    0.000    0.000    0.000 {method 'blit' of 'pygame.Surface' objects}
        1    0.000    0.000    0.000    0.000 {method 'disable' of '_lsprof.Profiler' objects}
        1    0.002    0.002    0.002    0.002 {method 'fill' of 'pygame.Surface' objects}
       18    0.000    0.000    0.000    0.000 {method 'random_sample' of 'mtrand.RandomState' objects}
        2    0.000    0.000    0.000    0.000 {method 'release' of '_thread.lock' objects}
        3    0.000    0.000    0.000    0.000 {method 'render' of 'pygame.font.Font' objects}

如您所见，整个 update() 函数(模拟的核心)和 brain.tick() 似乎要快得多。那为什么程序运行的时候比较慢呢？

干杯。

Answer 1

在您的新实现中，您为每个 Brain 对象创建了 5 个 numpy 数组:

self.h_activation = zeros((self.hidden_num, 1), dtype=float)
self.o_activation = zeros((self.output_num, 1), dtype=float)

self.i_output = zeros((self.input_num, 1), dtype=float)      
self.h_output = zeros((self.hidden_num, 1), dtype=float)  
self.o_output = zeros((self.output_num, 1), dtype=float)

这些属性在代码的其他部分没有被引用。创建它们是一项可能代价高昂的操作，在原始实现中似乎没有直接对应的操作。我不确定它是否会超过更快的 numpy 计算的速度优势，但如果您正在创建大量 Brain 对象，那么它值得一看。