Introduction

This repo contains a LSTM tensorflow model along with its deeplearn.js usage example to load and display in a browser a chunk of 512 characters every 10 secs.

The training data was taken from a compilation of poems written by Mihai Eminescu, a romanian poet. You can find the original data here.

Python sources can be found in src/python while the web code can be inspected in src/web.

It was presented at the Timisoara Deep Learning Meetup on LSTMs on 1 Februrary 2018. Slides can be accessed here.

Prerequisites

A Linux distro is required having the following tools installed:

• Python
• Tensorflow library (GPU vaiant preferably)
• NodeJS
• Yarn package manager

Usage

First you have to install the dependencies. You do this with:

yarn prep

Then you need to train the network. You can tweak the hyperparameters from src/python/params.py. To start the training, you type:

yarn train

Next you need to export the network's learned parameters along with compiling and deploying the web typescript sources (src/web/main.ts). You do this by typing:

yarn deploy

Having this done, you can see the deployed files in build folder.

To start serving the page and have it viewable from the browser, just type:

yarn start-server

wait 10 seconds and you shall see the first batch of text generated.

You can also do a yarn clean which deletes the build folder along with all the node_modules leaving you with a clean source base. Keep in mind that if you do this, you will have to redo all the previous steps to access the page again.

License

You are free to use the info/code you see here however you want with no restrictions.

Happy learning! 🍺