Keras 在 fit_generator 之后停止，没有错误

Question

我正在使用 Resnet 架构来比较性能与 CNN。我在第一次测试中使用了这个 Resnet。我正在重用我的代码来为网络加载和准备数据，它工作得很好。脚本的其余部分似乎也可以正常工作，除非它到达 fit_generator。在 fit_generator 它暂停了一段时间然后似乎退出了我有一个打印语句说“发生了什么？”我很困惑，因为我希望出现错误消息或程序崩溃或其他东西。我正在使用运行最新版本 anaconda 的 Windows 10。在我的condo环境中，我使用的是python 3.6，最新版本的Keras 2.3，最新版本的TensorFlow。如果有任何见解，我将不胜感激。

def batch_generator(X_train, Y_train):  
    while True:
        for fl, lb in zip(X_train, Y_train):
            sam, lam = get_IQsamples(fl, lb)
            max_iter = sam.shape[0]
            sample = []     # store all the generated data batches
            label = []   # store all the generated label batches

            i = 0
            for d, l in zip(sam, lam):
                sample.append(d)
                label.append(l)
                i += 1
                if i == max_iter:
                    break
            sample = np.asarray(sample)        
            label = np.asarray(label)
            yield sample, label


def residual_stack(x, f):
    
    # 1x1 conv linear
    x = Conv2D(f, (1, 1), strides=1, padding='same', data_format='channels_last')(x)
    x = Activation('linear')(x)


    # residual unit 1    
    x_shortcut = x
    x = Conv2D(f, (3, 2), strides=1, padding="same", data_format='channels_last')(x)
    x = Activation('relu')(x)
    x = Conv2D(f, 3, strides=1, padding="same", data_format='channels_last')(x)
    x = Activation('linear')(x)

    # add skip connection
    if x.shape[1:] == x_shortcut.shape[1:]:
      x = Add()([x, x_shortcut])

    else:
      raise Exception('Skip Connection Failure!')


    # residual unit 2    
    x_shortcut = x
    x = Conv2D(f, 3, strides=1, padding="same", data_format='channels_last')(x)
    x = Activation('relu')(x)
    x = Conv2D(f, 3, strides = 1, padding = "same", data_format='channels_last')(x)
    x = Activation('linear')(x)

    # add skip connection
    if x.shape[1:] == x_shortcut.shape[1:]:
      x = Add()([x, x_shortcut])

    else:
      raise Exception('Skip Connection Failure!')


    # max pooling layer
    x = MaxPooling2D(pool_size=2, strides=None, padding='valid', data_format='channels_last')(x)

    return x

.

定义 ResNet 模型

# define resnet model

def ResNet(input_shape, classes):   

    # create input tensor
    x_input = Input(input_shape)
    x = x_input

    # residual stack
    num_filters = 40
    x = residual_stack(x, num_filters)
    x = residual_stack(x, num_filters)
    x = residual_stack(x, num_filters)
    x = residual_stack(x, num_filters)
    x = residual_stack(x, num_filters)


    # output layer
    x = Flatten()(x)
    x = Dense(128, activation="selu", kernel_initializer="he_normal")(x)
    x = Dropout(.5)(x)
    x = Dense(128, activation="selu", kernel_initializer="he_normal")(x)
    x = Dropout(.5)(x)
    x = Dense(classes , activation='softmax', kernel_initializer = glorot_uniform(seed=0))(x)


    # Create model
    model = Model(inputs = x_input, outputs = x)
    model.summary()

    return model


model = ResNet((32,32,2),8)

model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])


print('Load complete!')
print('\n')


steps = val_length_train // batchsize
valid_steps = val_length // batchsize

history = model.fit_generator(
            generator=train_gen,
            epochs=3,
            verbose=0,
            steps_per_epoch=steps,
            validation_data=valid_gen,
            validation_steps=valid_steps,
            callbacks=[tensorboard])

print("what happened?")

Answer 1

有点。如果出现错误，它仍然会被抛出并打印详细信息为 0。也就是说，详细信息 0 似乎会给某些人带来问题。这篇文章来自 2017 年，但我看到了与 2019 年 11 月 https://github.com/keras-team/keras/issues/5818 相同的问题。如果我使用 0 或 2 项工作正常，但所有这些都无关紧要，因为脚本似乎从未开始抓取数据或训练。我很感激反馈。