使用MaxPool图层时Jupyter Notebook被杀死

Question

我正在使用Keras-CPU训练神经网络模型。当我不添加任何MaxPool层时，它运行良好，但是每当我使用MaxPool层时，我的jupyter内核就会死掉。

我的整个数据大小约为 480 MB。我的系统具有128GB的内存，并且在没有MaxPool层的情况下运行时不会占用超过5-10 GB的内存。我的批次大小也很小（32），并且运行培训时内存使用没有显着增加。

CPU使用率上升几秒钟，然后下降。关于单个时期的信息，有一个附加的度量；它只是打印时期的开始：

  

第1/2集

然后，我的内核死了。

这是网络的体系结构：

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, 382)               0         
_________________________________________________________________
embedding_1 (Embedding)      (None, 382, 100)          490600    
_________________________________________________________________
conv1d_1 (Conv1D)            (None, 382, 64)           12864     
_________________________________________________________________
max_pooling1d_1 (MaxPooling1 (None, 191, 64)           0         
_________________________________________________________________
conv1d_2 (Conv1D)            (None, 191, 128)          16512     
_________________________________________________________________
max_pooling1d_2 (MaxPooling1 (None, 95, 128)           0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 12160)             0         
_________________________________________________________________
dense_1 (Dense)              (None, 64)                778304    
_________________________________________________________________
dropout_1 (Dropout)          (None, 64)                0         
_________________________________________________________________
dense_2 (Dense)              (None, 128)               8320      
_________________________________________________________________
dense_3 (Dense)              (None, 1)                 129       
=================================================================
Total params: 1,306,729
Trainable params: 1,306,729
Non-trainable params: 0

这是我训练神经网络的方法：

def get_nnn(train_data_shape, max_val, emb_shape, drop_out, pre_batch_size, pre_batch_epoch,\
           final_batch_size, final_batch_epoch, ls_start, lr_end,class_weight, lr_decay_pat,\
                      lr_decay_rat,num_conv,conv_kernels,conv_neurons,train_data,                              train_lbl,test_data,test_lbl,loss_type='binary_crossentropy',\
           act='sigmoid',use_emb=True,use_max_pool=True,custom_methods=[]):

    #A lot of params

    nnet = Input(shape=(train_data_shape,))
    inp = nnet

    if use_emb:
        nnet = Embedding(max_val+1, emb_shape)(nnet)

    for i in range(num_conv):
        nnet = Conv1D(conv_neurons[i],conv_kernels[i], padding='same',activation='relu')(nnet)
        if use_max_pool:
            #Adding MaxPool here
            nnet = MaxPooling1D(pool_size=2)(nnet)
    nnet = Flatten()(nnet)
    nnet = Dense(64, activation='relu')(nnet)
    nnet = Dropout(drop_out)(nnet)
    nnet = Dense(128, activation='relu')(nnet)

    batch_size = final_batch_size
    epochs = final_batch_epoch

    lentrain = train_data.shape[0]

    exp_decay = lambda init, fin, steps: (init/fin)**(1/(steps-1)) - 1
    steps = int(lentrain / batch_size) * epochs
    lr_init, lr_fin = ls_start, lr_end
    lr_decay = exp_decay(lr_init, lr_fin, steps)

    output = Dense(1, activation=act)(nnet)
    model = Model(inputs=inp, outputs=output)


    model.compile(loss=loss_type, optimizer=Adam(lr=ls_start, decay=lr_decay),
                  metrics=['accuracy',precisioned,recalled,f1]+custom_methods)


    reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=lr_decay_rat,
                              patience=lr_decay_pat, min_lr=lr_end)


    print(model.summary())

    model.fit(train_data, train_lbl,
              epochs=pre_batch_epoch,
              batch_size=pre_batch_size,
              class_weight=class_weight,
              validation_data=(test_data, test_lbl),
              callbacks=[reduce_lr])

    model.fit(train_data, train_lbl,
              epochs=epochs,
              batch_size=batch_size,
              class_weight=class_weight,
              validation_data=(test_data, test_lbl),
              callbacks=[reduce_lr])

    return model