BootstrapNLP

BootstrapNLP is a Python-based web server that will serve an easy to interact with webpage which will push and pull directly to a pre-trained tensorflow model.

The intent of this application is to enable ML developers to quickly get a freshly trained model into a UI that can be used for testing or demonstration purposes. There are no fancy pipelines or architectures to learn. You can be up and running in a couple of minutes.

The application is primarily configuration driven. The web page it serves can host multiple models and allows you to configure multiple text inputs and outputs.

This project is still very much a work in progress, but here are some tips in case you want to give it a try:

Getting Started

You will bring:

1. A trained tensorflow model that you will serve using tensorflow serving
2. An input and output processor which will take in text and produce inputs for your model and take outputs and convert them into something pretty to view on the webpage. Generally, the input processor is simply your tokenizer.

To get this thing running, here is what you will do:

1. Start up a tensorflow_serving server. Here is the command I use:

docker run -p 8500:8500 --mount type=bind,source=<saved_model_path>,target=<model_name_inside_saved_path> -e MODEL_NAME=<model_id, ex: 'gpt2'> -t tensorflow/serving

2. Edit model_configs.yaml from this repo. I've extensively documented this configuration file. You will need to provide some basic information about your model, the inputs it takes, the outputs it produces, and where the TF model server resides. Use the values I provided as a guide.
3. Create your data processing script which was specified in the config. Generally the inputs will just perform basic tokenization and the outputs will be fed directly through. See processors/sentiment.py for an example.
4. Run flask_server.py and point your browser to http://localhost:8080

Future Changes

I have a few focuses for this project at the moment:

1. Move from tensorflow serving to a model serving framework that supports pytorch as well as tensorflow. I'm looking closely at RedisAI for this right now.
2. Retool HTML page to use Angular or React rather than a custom data binding script.
3. Add additional input and output types as needed.

I will endeavor to only commit working code to GitHub. Pull requests are welcome.