network.py 文件源码

def build(features_shape, audio_spectrogram_size):
        model = Sequential()

        model.add(Flatten(input_shape=features_shape))
        model.add(BatchNormalization())

        model.add(Dense(1024, kernel_initializer='he_normal', name='dense1'))
        model.add(BatchNormalization())
        model.add(LeakyReLU())
        model.add(Dropout(0.25))

        model.add(Dense(1024, kernel_initializer='he_normal', name='dense2'))
        model.add(BatchNormalization())
        model.add(LeakyReLU())
        model.add(Dropout(0.25))

        model.add(Dense(1024, kernel_initializer='he_normal', name='dense3'))
        model.add(BatchNormalization())
        model.add(LeakyReLU())
        model.add(Dropout(0.25))

        model.add(Dense(audio_spectrogram_size, name='output'))

        model.summary()

        return VideoToSpeechNet(model)