Character Level LSTM - Tensorflow/MLflow

Temporary Url: http://ec2-3-94-194-50.compute-1.amazonaws.com/

This repository contains an Mlflow deployment of my character-level LSTM. Since the LSTM executes custom logic when evaluating queries, it is represented as an mlflow.pyfunc model.

Installation

1. Clone repo locally
2. Install dependencies with pip3 install -r requirements.txt
3. Train model from scratch or download saves from the link below into the saves folder
4. Run python3 create_model_wrapper.py to create pyfunc model in /model_path
5. Run python3 application.py to start the flask server

Architecture

I trained a standard multi-layer cahracter level LSTM with the following cell archictecture. This project was heavily inspired by Andrej Karpathy's blog post http://karpathy.github.io/2015/05/21/rnn-effectiveness/

The network has three stacked layers of cells with a batch sequence length size of 100. I trained on an single NVIDIA GeForce 1080.

Download pre-trained model here: https://www.dropbox.com/s/o0glxwlbwtesdck/saves.zip?dl=0

Random weights:

U—CN3gLBddDWÁu!HÁ”;ap/8!ôdPEç’hÁóblOxD/ç)Vsn@i)Á”YáPxwDöAhDœ “uwDóJ$—u)cDèBDü->èAàZ@zàB(ióH$äaFu3ýYmFniRLMÉKmYwBwZO@Uuaîq@AG()%aèLî$5b%üpïER=—‘À*mê?ç:V v B’BlcFIy5hf98JouV jAúoöt(y(WWnRE;él:xGlêEWÀnoG

Here are some examples of text halucinations:

the ball, calmls happened to you, does he cavered in the steps to maken, narged the wortured loud. The doctor and from his ?? fatare and ran now a run into the existence of the great some! They were saying, he felt a drunk as you weight he saw in its away you on leaving a reflection”arch, and all they had till the army and are no waiting those people was heart free terringly with somewhere.

“To pass? Dudly—retelies to them said for his place of Austria consulting impossible, more still told his corrigencial it had way out of honors could not been glanced lea into the comrades merely to a relate.

Without will be a the people she told not your house, looking began unterestby plump ask.

The old fair that that it he has escort of our dresses ready, but his dwant to you wagrag immediately if bridge and rouding to kiss a whit of relay of, he must do so, is not imaginex regard to really singers, and exembrowed smiles, stapid vicious own ind-Ardisoners, quition, and I have letter she rose and