训练使用 tensorflow.keras.Model 和 keras 函数式 API 设计的网络会导致 Python 崩溃

Question

-Keras 版本：2.3.1 -Tensorflow 版本：2.2.0 -OS：Windows 10 -运行于：CPU 处理器 Intel(R) Core(TM) i7-8850H -Developed in：PyCharm -Python 版本：3.7

我曾尝试训练使用 keras 函数式 API 设计的网络，但是训练总是导致 python 崩溃并显示以下退出消息：

Process finished with exit code -1073740791 (0xC0000409)

不提供其他堆栈跟踪。 下面是我试图运行的代码的简化版本：

import scipy.io as io
import os
import numpy as np
from tensorflow.keras import layers, losses, Input
from tensorflow.keras.models import Model

directory = 'C:\\Dataset'

def simple_generator(dim1, dim2, dim3, batch_size=5):
    while True:
        samples = np.random.random_sample((batch_size, dim1, dim2, dim3))
        targets = np.random.random_sample((batch_size, 3))
        yield samples, targets

example_shape = (1100, 4096, 2)

train_gen = simple_generator(example_shape[0], example_shape[1], example_shape[2], batch_size=10)
val_gen = simple_generator(example_shape[0], example_shape[1], example_shape[2])
test_gen = simple_generator(example_shape[0], example_shape[1], example_shape[2])

input_tensor = Input(shape=example_shape)
preprocess = layers.Conv2D(32, 3, activation='relu')(input_tensor)
preprocess = layers.Conv2D(32, 3, activation='relu')(preprocess)
preprocess = layers.MaxPool2D(pool_size=(3, 3), strides=3)(preprocess)

"""   Head 1   """
head_1 = layers.Conv2D(32, 3, activation='relu')(preprocess)
head_1 = layers.Conv2D(32, 3, activation='relu')(head_1)
head_1 = layers.MaxPool2D(pool_size=(3, 3), strides=3)(head_1)
head_1 = layers.Conv2D(32, 3, activation='relu')(head_1)
head_1 = layers.Conv2D(32, 3, activation='relu')(head_1)
head_1 = layers.MaxPool2D(pool_size=(3, 3), strides=3)(head_1)
head_1 = layers.Conv2D(16, 3, activation='relu')(head_1)
head_1 = layers.Conv2D(16, 3, activation='relu')(head_1)
head_1 = layers.MaxPool2D(pool_size=(3, 3), strides=3)(head_1)
head_1 = layers.Dense(8, activation='relu')(head_1)

"""   Head 2   """
head_2 = layers.Conv2D(32, 3, activation='relu')(preprocess)
head_2 = layers.Conv2D(32, 3, activation='relu')(head_2)
head_2 = layers.MaxPool2D(pool_size=(3, 3), strides=3)(head_2)
head_2 = layers.Conv2D(32, 3, activation='relu')(head_2)
head_2 = layers.Conv2D(32, 3, activation='relu')(head_2)
head_2 = layers.MaxPool2D(pool_size=(3, 3), strides=3)(head_2)
head_2 = layers.Conv2D(16, 3, activation='relu')(head_2)
head_2 = layers.Conv2D(16, 3, activation='relu')(head_2)
head_2 = layers.MaxPool2D(pool_size=(3, 3), strides=3)(head_2)
head_2 = layers.Dense(8, activation='relu')(head_2)

"""   Head 3   """
head_3 = layers.Conv2D(32, 3, activation='relu')(preprocess)
head_3 = layers.Conv2D(32, 3, activation='relu')(head_3)
head_3 = layers.MaxPool2D(pool_size=(3, 3), strides=3)(head_3)
head_3 = layers.Conv2D(32, 3, activation='relu')(head_3)
head_3 = layers.Conv2D(32, 3, activation='relu')(head_3)
head_3 = layers.MaxPool2D(pool_size=(3, 3), strides=3)(head_3)
head_3 = layers.Conv2D(16, 3, activation='relu')(head_3)
head_3 = layers.Conv2D(16, 3, activation='relu')(head_3)
head_3 = layers.MaxPool2D(pool_size=(3, 3), strides=3)(head_3)
head_3 = layers.Dense(8, activation='relu')(head_3)

concat_out = layers.Concatenate(axis=-1)([head_1, head_2, head_3])
concat_out = layers.Flatten()(concat_out)
output_tensor = layers.Dense(3)(concat_out)
model = Model(input_tensor, output_tensor)

model.compile(loss=losses.mean_absolute_error, optimizer='sgd')
model.summary()

history = model.fit(x=train_gen, steps_per_epoch=25, epochs=12, validation_data=val_gen, validation_steps=10, verbose=True)
model.save(directory + '\\divergent_network.h5')
evaluations = model.evaluate_generator(generator=test_gen, steps=20, verbose=True)
print(evaluations)

任何人都可以解释崩溃或提出任何潜在的解决方案吗？ 我在 Anaconda 环境中运行此代码，但是我尝试在基本的 python 虚拟环境中运行以查看它是否是 Anaconda 问题，但我发现两者之间没有区别。

Answer 1

我的 conda 环境似乎有问题。 我创建了一个新的“最低限度”环境，其中只包含我需要的软件包，并清除了所有内容。 我用来执行此操作的信息可以在以下位置找到：

https://docs.conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html#creating-an-environment-with-commands

https://www.jetbrains.com/help/pycharm/creating-virtual-environment.html