CaptionModel.py

import keras

def create_model(max_caption_length, vocab_length):
    
    # sub network for handling the image feature part
    input_layer1 = keras.Input(shape=(18432))
    feature1 = keras.layers.Dropout(0.2)(input_layer1)
    feature2 = keras.layers.Dense(max_caption_length*4, activation='relu')(feature1)
    feature3 = keras.layers.Dense(max_caption_length*4, activation='relu')(feature2)
    feature4 = keras.layers.Dense(max_caption_length*4, activation='relu')(feature3)
    feature5 = keras.layers.Dense(max_caption_length*4, activation='relu')(feature4)
    
    # sub network for handling the text generation part
    input_layer2 = keras.Input(shape=(max_caption_length,))
    cap_layer1 = keras.layers.Embedding(vocab_length, 300, input_length=max_caption_length)(input_layer2)
    cap_layer2 = keras.layers.Dropout(0.2)(cap_layer1)
    cap_layer3 = keras.layers.LSTM(max_caption_length*4, activation='relu', return_sequences=True)(cap_layer2)
    cap_layer4 = keras.layers.LSTM(max_caption_length*4, activation='relu', return_sequences=True)(cap_layer3)
    cap_layer5 = keras.layers.LSTM(max_caption_length*4, activation='relu', return_sequences=True)(cap_layer4)
    cap_layer6 = keras.layers.LSTM(max_caption_length*4, activation='relu')(cap_layer5)
    
    # merging the two sub network
    decoder1 = keras.layers.merge.add([feature5, cap_layer6])
    decoder2 = keras.layers.Dense(256, activation='relu')(decoder1)
    decoder3 = keras.layers.Dense(256, activation='relu')(decoder2)
    
    # output is the next word in sequence
    output_layer = keras.layers.Dense(vocab_length, activation='softmax')(decoder3)
    model = keras.models.Model(inputs=[input_layer1, input_layer2], outputs=output_layer)
    
    model.summary()

    return model