model.py 文件源码

def create_model(config_dict,
                 compile_model=True):
    image_inputs = Input(shape=(4096,), name="image_model_input")
    image_model = _create_image_model(config_dict=config_dict,
                                      image_inputs=image_inputs)

    language_inputs = Input(shape=(config_dict['max_caption_length'],),
                            name="language_model_input")
    language_model = _create_language_model(config_dict=config_dict,
                                            language_inputs=language_inputs)

    merged_input = concatenate([image_model, language_model],
                               name="concatenate_image_language")
    merged_input = LSTM(1000,
                        return_sequences=False,
                        name="merged_model_lstm")(merged_input)
    softmax_output = Dense(units=config_dict["vocabulary_size"],
                           activation="softmax",
                           name="merged_model_softmax")(merged_input)
    model = Model(inputs=[image_inputs,
                          language_inputs], outputs=softmax_output)
    print(model.summary())
    if (compile_model == True):
        model.compile(loss='categorical_crossentropy',
                      optimizer='adam',
                      metrics=['accuracy'])
    return model