使用Keras在冷杉神经网络中精度降低

Question

我有这个实现神经网络的功能：

    def create(self, size_1l, size_2l, size_3l=0, size_4l=0):
    """
    This function build the denoising autoencoder
    using the parameters as size, the first is the input layer,
    the third and the fourth if equal to 0 are omitted.
    """
    input_layer = Input(shape=(size_1l, size_1l, 1))
    # encoder
    x = Conv2D(size_2l, (3, 3), activation='relu', padding='same', use_bias=True)(input_layer)
    if size_3l > 0:
        x = MaxPooling2D((2, 2), padding='same')(x)
        x = Conv2D(size_3l, (3, 3), activation='relu', padding='same', use_bias=True)(x)
    if size_4l > 0:
        x = MaxPooling2D((2, 2), padding='same')(x)
        x = Conv2D(size_4l, (3, 3), activation='relu', padding='same', use_bias=True)(x)
    x = MaxPooling2D((2, 2), padding='same')(x)  # encoded layer
    # decoder
    if size_4l > 0:
        x = Conv2D(size_4l, (3, 3), activation='relu', padding='same', use_bias=True)(x)
        x = UpSampling2D((2, 2))(x)
    if size_3l > 0:
        x = Conv2D(size_3l, (3, 3), activation='relu', padding='same', use_bias=True)(x)
        x = UpSampling2D((2, 2))(x)
    x = Conv2D(size_2l, (3, 3), activation='relu', padding='same', use_bias=True)(x)
    x = UpSampling2D((2, 2))(x)
    output_layer = Conv2D(1, (3, 3), activation='tanh', padding='same', use_bias=True)(x)
    # building
    self.autoencoder = Model(input_layer, output_layer)
    self.autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy', metrics=['accuracy'])

这个培训代码：

    def train(self, training_data, target_data, training_test_data, target_test_data, epochs=100, batch_size=32, verbose=0):
    history = self.autoencoder.fit(training_data, target_data,
                                   epochs=epochs,
                                   batch_size=batch_size,
                                   shuffle=True,
                                   verbose=1,
                                   validation_data=(training_test_data, target_test_data))

问题是在训练期间准确度值会不断下降，如下所示：

有一些建议吗？怎么了？为什么在学习过程中准确性的价值不会增加？

由于