Question

我正在尝试使用keras在python2中提供的model.predict命令并行执行模型预测。我将tensorflow 1.14.0用于python2。我有5个模型（.h5）文件，并希望预测命令并行运行，这是在python 2.7中运行的。我正在使用多处理池，通过预测功能将模型文件名映射到多个进程，如下所示，

import matplotlib as plt
import numpy as np
import cv2
from multiprocessing import Pool
pool=Pool()
def prediction(model_name):
    global input
    from tensorflow.keras.models import load_model
    model=load_model(model_name)
    ret_val=model.predict(input).tolist()[0]
    return ret_val

models=['model1.h5','model2.h5','model3.h5','model4.h5','model5.h5']
start_time=time.time()
res=pool.map(prediction,models)
print('Total time taken: {}'.format(time.time() - start_time))
print(res)

输入是从代码另一部分获得的图像numpy数组。但是执行此操作后，我得到以下信息，

Traceback (most recent call last):
Traceback (most recent call last):
  File "/usr/lib/python2.7/multiprocessing/process.py", line 267, in _bootstrap
  File "/usr/lib/python2.7/multiprocessing/process.py", line 267, in _bootstrap
    self.run()
    self.run()
  File "/usr/lib/python2.7/multiprocessing/process.py", line 114, in run
    self._target(*self._args, **self._kwargs)
  File "/usr/lib/python2.7/multiprocessing/process.py", line 114, in run
  File "/usr/lib/python2.7/multiprocessing/pool.py", line 102, in worker
    self._target(*self._args, **self._kwargs)
  File "/usr/lib/python2.7/multiprocessing/pool.py", line 102, in worker
    task = get()
  File "/usr/lib/python2.7/multiprocessing/queues.py", line 376, in get
    task = get()
  File "/usr/lib/python2.7/multiprocessing/queues.py", line 376, in get
    return recv()
    return recv()
AttributeError: 'module' object has no attribute 'prediction'
AttributeError: 'module' object has no attribute 'prediction'

我无法解释此错误消息，我该如何解决呢？任何建议都非常感谢！

更新2： 感谢所有的指针和完整的示例@sokato。我执行了@sokato发布的确切代码，但是遇到了以下错误（我也对代码进行了更改，并得到如下所示的相同错误），

Traceback (most recent call last):
  File "stackoverflow.py", line 47, in <module>
    with multiprocessing.Pool() as p:
AttributeError: __exit__

UPDATE3： 感谢所有支持。我认为 UPDATE2 中的问题是由于使用了python2而不是python3。通过使用with closing(multiprocessing.Pool()) as p:而非@sokato代码中的with multiprocessing.Pool() as p:，我能够解决python2的 UPDATE2 中给出的错误。导入关闭函数，如下所示：from contextlib import closing

使用下面显示的不同方法的新问题，

实际上我有多个输入。我不想每次都为每个输入加载模型，而是想事先加载所有模型并将其保存在列表中。我已经做到了，如下所示，

import matplotlib as plt
import numpy as np
import cv2
import multiprocessing
import tensorflow as tf
from contextlib import closing
import time

models=['model1.h5','model2.h5','model3.h5','model4.h5','model5.h5']
loaded_models=[]
for model in models:
    loaded_models.append(tf.keras.models.load_model(model))

def prediction(input_tuple):
    inputs,loaded_models=input_tuple
    predops=[]
    for model in loaded_models:
        predops.append(model.predict(inputs).tolist()[0])
    actops=[]
    for predop in predops:
        actops.append(predop.index(max(predop)))
    max_freqq = max(set(actops), key = actops.count) 
    return max_freqq

#....some pre-processing....#

    '''new_all_t is a list which contains tuples and each tuple has inputs from all_t 
    and the list containing loaded models which will be extracted
 in the prediction function.'''

new_all_t=[]
for elem in all_t:
    new_all_t.append((elem,loaded_models))
start_time=time.time()
with closing(multiprocessing.Pool()) as p:
    predops=p.map(prediction,new_all_t)
print('Total time taken: {}'.format(time.time() - start_time))

new_all_t是一个包含元组的列表，每个元组都有来自all_t的输入，该列表包含将在预测函数中提取的已加载模型。但是，我现在遇到以下错误，

Traceback (most recent call last):
  File "trial_mult-ips.py", line 240, in <module>
    predops=p.map(prediction,new_all_t)
  File "/usr/lib/python2.7/multiprocessing/pool.py", line 253, in map
    return self.map_async(func, iterable, chunksize).get()
  File "/usr/lib/python2.7/multiprocessing/pool.py", line 572, in get
    raise self._value
NotImplementedError: numpy() is only available when eager execution is enabled.

这究竟表明了什么？我该如何解决这个问题？

UPDATE4： 我加入了tf.compat.v1.enable_eager_execution()行， tf.compat.v1.enable_v2_behavior()一开始。现在我得到以下错误，

WARNING:tensorflow:From /home/nick/.local/lib/python2.7/site-packages/tensorflow/python/ops/math_grad.py:1250: where (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Use tf.where in 2.0, which has the same broadcast rule as np.where

Traceback (most recent call last):
  File "the_other_end-mp.py", line 216, in <module>
    predops=p.map(prediction,modelon)
  File "/usr/lib/python2.7/multiprocessing/pool.py", line 253, in map
    return self.map_async(func, iterable, chunksize).get()
  File "/usr/lib/python2.7/multiprocessing/pool.py", line 572, in get
    raise self._value
ValueError: Resource handles are not convertible to numpy.

我无法解释此错误消息，我该如何解决呢？任何建议都非常感谢！

Answer 1

因此，我不确定您的某些设计选择，但我已根据给定的信息进行了最佳尝试。具体来说，我认为并行函数中的全局变量和import语句可能存在一些问题。

您应该使用共享变量而不是全局变量在进程之间共享输入。如果需要，可以在多处理文档中阅读有关共享内存的更多信息。
由于不包含您的模型，因此我从教程中生成了模型。
您没有加入或关闭池，但是通过以下代码，我能够使代码成功并行执行。您可以通过调用pool.close()或使用下面显示的“ with”语法来关闭池。请注意，with语法不适用于python 2.7。

import numpy as np
import multiprocessing, time, ctypes, os
import tensorflow as tf

mis = (28, 28) #model input shape
mnist = tf.keras.datasets.mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0

def createModels(models):
    model = tf.keras.models.Sequential([
        tf.keras.layers.Flatten(input_shape=mis),
        tf.keras.layers.Dense(128, activation='relu'),
        tf.keras.layers.Dropout(0.2),
        tf.keras.layers.Dense(10)
    ])

    model.compile(optimizer='adam',
               loss=tf.losses.SparseCategoricalCrossentropy(from_logits=True),
               metrics=['accuracy'])

    model.fit(x_train, y_train, epochs=5)

    for mod in models:
        model.save(mod)

def prediction(model_name):

    model=tf.keras.models.load_model(model_name)
    ret_val=model.predict(input).tolist()[0]
    return ret_val

if __name__ == "__main__":
    models=['model1.h5','model2.h5','model3.h5','model4.h5','model5.h5']
    dir = os.listdir(".")
    if models[0] not in dir:
        createModels(models)
    # Shared array input
    ub = 100
    testShape = x_train[:ub].shape
    input_base = multiprocessing.Array(ctypes.c_double, 
    int(np.prod(testShape)),lock=False)
    input = np.ctypeslib.as_array(input_base)
    input = input.reshape(testShape)
    input[:ub] = x_train[:ub]

    # with multiprocessing.Pool() as p:  #Use me for python 3
    p = multiprocessing.Pool() #Use me for python 2.7
    start_time=time.time()
    res=p.map(prediction,models)
    p.close() #Use me for python 2.7
    print('Total time taken: {}'.format(time.time() - start_time))
    print(res)

我希望这会有所帮助。

使用针对Python的多处理功能在keras中并行化模型预测

1 个答案: