utils.py 文件源码

def create_imgText(image_shape, max_caption_len,vocab_size):
    image_model = Sequential()
    # image_shape : C,W,H
    # input: 100x100 images with 3 channels -> (3, 100, 100) tensors.
    # this applies 32 convolution filters of size 3x3 each.
    image_model.add(Convolution2D(32, 3, 3, border_mode='valid', input_shape=image_shape))
    image_model.add(BatchNormalization())
    image_model.add(Activation('relu'))
    image_model.add(Convolution2D(32, 3, 3))
    image_model.add(BatchNormalization())
    image_model.add(Activation('relu'))
    image_model.add(MaxPooling2D(pool_size=(2, 2)))
    image_model.add(Dropout(0.25))
    image_model.add(Convolution2D(64, 3, 3, border_mode='valid'))
    image_model.add(BatchNormalization())
    image_model.add(Activation('relu'))
    image_model.add(Convolution2D(64, 3, 3))
    image_model.add(BatchNormalization())
    image_model.add(Activation('relu'))
    image_model.add(MaxPooling2D(pool_size=(2, 2)))
    image_model.add(Dropout(0.25))
    image_model.add(Flatten())
    # Note: Keras does automatic shape inference.
    image_model.add(Dense(128))
    image_model.add(RepeatVector(1))
    #model = AttentionSeq2Seq(input_dim=128, input_length=1, hidden_dim=128, output_length=max_caption_len, output_dim=vocab_size)
    model = Seq2Seq(input_dim=128, input_length=1, hidden_dim=128, output_length=max_caption_len,
                             output_dim=128, peek=True)
    image_model.add(model)
    image_model.add(TimeDistributed(Dense(vocab_size)))
    image_model.add(Activation('softmax'))
    return image_model