Keras(TensorFlow,CPU):训练循环中的顺序模型吃内存

Question

我试图在循环中训练1000x的Sequential模型.在每个循环中,我的程序都会泄漏内存,直到我用完并获得OOM异常.

之前我已经问了一个类似的问题(连续训练多个连续模型减速)

并且看到其他类似的问题(Keras:进行超参数网格搜索时内存不足)

并且在K.clear_session()使用完模型后,解决方案始终是添加到代码中.所以我在上一个问题中做到了这一点,我仍在泄露记忆

这是重现问题的代码.

import random
import time
from keras.models import Sequential
from keras.layers import Dense
from keras import backend as K
import tracemalloc


def run():
    tracemalloc.start()
    num_input_nodes = 12
    num_hidden_nodes = 8
    num_output_nodes = 1

    random_numbers = random.sample(range(1000), 50)
    train_x, train_y = create_training_dataset(random_numbers, num_input_nodes)

    for i in range(100):
        snapshot = tracemalloc.take_snapshot()
        for j in range(10):
            start_time = time.time()
            nn = Sequential()
            nn.add(Dense(num_hidden_nodes, input_dim=num_input_nodes, activation='relu'))
            nn.add(Dense(num_output_nodes))
            nn.compile(loss='mean_squared_error', optimizer='adam')
            nn.fit(train_x, train_y, nb_epoch=300, batch_size=2, verbose=0)
            K.clear_session()
            print("Iteration {iter}. Current time {t}. Took {elapsed} seconds".
                  format(iter=i*10 + j + 1, t=time.strftime('%H:%M:%S'), elapsed=int(time.time() - start_time)))

        top_stats = tracemalloc.take_snapshot().compare_to(snapshot, 'lineno')

        print("[ Top 5 differences ]")
        for stat in top_stats[:5]:
            print(stat)


def create_training_dataset(dataset, input_nodes):
    """
    Outputs a training dataset (train_x, train_y) as numpy arrays.
    Each item in train_x has 'input_nodes' number of items while train_y items are of size 1
    :param dataset: list of ints
    :param input_nodes:
    :return: (numpy array, numpy array), train_x, train_y
    """
    data_x, data_y = [], []
    for i in range(len(dataset) - input_nodes - 1):
        a = dataset[i:(i + input_nodes)]
        data_x.append(a)
        data_y.append(dataset[i + input_nodes])
    return numpy.array(data_x), numpy.array(data_y)

run()

这是我从第一个内存调试打印得到的输出

/tensorflow/python/framework/ops.py:121:size = 3485 KiB(+3485 KiB),count = 42343(+42343)/tensorflow/python/framework/ops.py:1400:size = 998 KiB(+998 KiB),count = 8413(+8413)/tensorflow/python/framework/ops.py:116:size = 888 KiB(+888 KiB),count = 32468(+32468)/tensorflow/python/framework/ops.py :1185:size = 795 KiB(+ 795 KiB),count = 3179(+3179)/tensorflow/python/framework/ops.py:2354:size = 599 KiB(+599 KiB),count = 5886(+5886)

系统信息:

• python 3.5
• keras(1.2.2)
• tensorflow(1.0.0)

Answer 1

内存泄漏源于Keras和TensorFlow,使用单个"默认图形"来存储网络结构,随着内部for循环的每次迭代,网络结构的大小都会增加.

调用K.clear_session()在迭代之间释放与默认图关联的一些(后端)状态,但需要额外调用tf.reset_default_graph()以清除Python状态.

请注意,可能有一个更有效的解决方案:因为nn不依赖于任何一个循环变量,您可以在循环外定义它,并在循环内重用相同的实例.如果这样做,则无需清除会话或重置默认图表,并且性能会提高,因为您可以从迭代之间的缓存中受益.