Practice in Caption Generation with Keras: The Design and Evaluation for Attention Models
Practice in Caption Generation with Keras: The Design and Evaluation for Attention Models
复制标题
Keras 字幕生成实践：注意力模型的设计与评估
DOI：
10.1145/3342999.3343004
复制
发表时间：
2019
期刊：
影响因子：
--
通讯作者：
Toru Wakahara
中科院分区：
文献类型：
--
作者：
Rong Wang;Toru Wakahara
硬注意模型已被提出用于自动生成图像的字幕。目前，这种模式以多种方式实现。例如，它的原始实现是用Theano编写的，而大多数其他实现都是使用Tensorflow库编写的。在本文中，我们使用Keras，一个用于构建深度神经网络的模块化框架，来实现这个模型。我们还提出了两个注意力模型，称为Attention-L和Attention-C，从原来的注意力模型略有修改。使用基准数据集Flickr 30对这三种注意力模型进行了比较，结果表明Attention-C模型比其他两种模型更有可能获得更好的分数。
Hard-attention model has been proposed for automatically generating captions for images. Currently this model is implemented in many ways. For instance, its original implementation is written in Theano and most other implementations are written using Tensorflow library. In this paper, we use Keras, a modular framework for constructing deep neural networks, to implement this model. We also propose two attention models, called Attention-L and Attention-C, are slightly modified from the original attention model. The benchmark dataset Flickr30 is used to compare these three attention models, and the results demonstrate Attention-C model is more likely to obtain the better scores than that of other two models.