【bug记录】python内存泄漏-CSDN博客

本文链接：https://blog.csdn.net/weixin_42486139/article/details/107911479

问题描述：

多组数据需要多个模型，循环训练神经网络，结果发现程序越跑越慢

解决方法：

使用memory_profiler模块检测内存。安装： pip install memory_profiler

from memory_profiler import profile

@profile
def Conv_LSTM_2D( n_seq, n_steps, n_features):
    model = Sequential()
    model.add(ConvLSTM2D(filters = 64, kernel_size = (1, 2), padding='same', activation = 'relu',return_sequences = True,
                         input_shape = ( n_seq, 1, n_steps, n_features)))
    model.add(BatchNormalization())
    model.add(ConvLSTM2D(filters = 64, kernel_size = (1, 2),  activation = 'relu'))
    model.add(Flatten())
    model.add(Dense(1))
    model.compile(optimizer = 'adam', loss = 'mse')
    #print(model.summary())
    return model

for i in range(len(data)):
    #write_list = []
    info = data[i,17:75]
    #print(info)
    n_seq = 2
    n_steps = 4
    num_features = 1
    X, y= split_sequence(info, n_steps)
    model = Conv_LSTM_2D( 2, 2, 1)
    X = X.reshape((X.shape[0], n_seq, 1, n_steps, num_features))   #for Conv LSTM
    model.fit(X, y, epochs = 500, verbose = 0)

使用profile装饰器监测每次调用Conv_LSTM_2D函数时的内存占用，会发现，随着迭代次数的增加，内存一直在增加，程序越跑越慢。这是因为把网络的创建放在循环里，每次循环时都创建了计算图，导致内存占用越来越大，最终程序崩溃。

把建立模型这句代码model = Conv_LSTM_2D( 2, 2, 1)，移到循环外，问题解决。