By vivek | August 25, 2017 | 0 Comment
In this project I have connected Google’s tensorflow seq2seq to Google home. So basically I have connected a deep learning chatbot with google home.
Tensorflow is one of the excellent libraries which can be used for deep learning. They have open sourced there Sequence to Sequence based deep learning library which works on RNN (Recurrent neural networks) or in simple terms “backward and forward propagation”. So basically it accepts a “sequence” of dataset and produces the “sequence” of output.
(Image Source : Here )
The results are not so good, but it is really fun if you can directly integrate a deep learning chatbot with google home. And teach the same. I am going to publish some of my works of self-learning in upcoming months once I have finished working on it. I have used the github modified library by Nicolas Ivanov who has done really great work in conversation deep learning chatbot.
Initial library provided by Google for seq2seq, they have given the example of language translator. Which is used to translate language from one language to another. And it works decent.
So basically, I will give input to the chatbot as sequence of question and answers in conversation format. And expectation is to generate answers from that dataset. It didn’t work well.
I managed to make it work and thought of connecting it Google home. It was real pain since API.ai(the conversational platform where we have to develop apps for Google home) won’t allow me to interact with a non-hosted server(chatbot server running on local system) directly. So Other way around was to get it through a middle ware app hosted on heroku. And from heroku I was calling my server which was basically a script getting me the conversation output of the question which I will feed in.
I already trained some of my dataset. And the result was astonishing. Atleast the gimmicks part was.