About

More stars - more commits! You can do it too, but you can't be bothered now, can you?

MODELWRAPPER

An API+SDK+Storage+Threading framework for running machine learning models.

Changelog

v1.0 Base upload

TO-DO

Rewrite README
Add versioning to API
Add GOLANG integration
Set logger to use ConfigLoader
- Also change logging level on the fly
API Auto-tests
Authentication
Dashboard
Model dump depth config
Microservice composition option
Service discovery
Make container lightweight (get rid of CentOS and do the abose microservice thing)
EVEN MORE

Out of the box

Run willdrug/modelwrapper, while mounting the appropriate volumes:

/models_handler/models
/models_handler/dumps
/redis/data Upload your models to /models_handler/models volume and either issue a restart or, if you have set autorestart, issue a DELETE method to /service endpoint.

Features

API

Default is Flask-based REST API (as shown in start.py). Provides a port into running tasks, changing current config and seeing stats. Validates inbound params.

Tasker

Default is Threaded, provides a way to register a function to be ran as a separate thread

Config Loader

Uploads provided config (extending base class) into cache, making it persistent. Refreshes online conf (like API port) upon restart.

ModelWrapper

Provides basic FIT and PREDICT proxies, which can be ran as tasks. You are welcome to write your own.

Models are uploaded as packages into /models dir (docker volume). The only requirement is that they extend base classes provided. Connectors are provided for most popular data sources. Each model uploaded is stored on the shelf and you can restore to the previous versions (keep an eye on data to avoid reaching the same undesirable result after the next fit!)

Credits and links

Docker image is available at dockerhub

Credit to Nario for logger and modelwrapper base.

WillDrug/modelwrapper