Beaver is a system to deploy and maintain machine learning models. In particular, it caters to online models, which includes inference as well learning on streaming data.
Beaver is the latest of an ongoing trial and error process. The compromise we've settled on is to answer 80% of needs without worrying too much about scale, performance, security, etc. We want to provide a piece of software which takes care of (most of) the boring stuff, is pluggable, and is fun to use.
Much of Beaver's design can be traced back to some of Max Halford's talks, with this one in particular.
The central object is the Dam. It holds all the components necessary to perform online machine learning. For instance, a data_store which will store predictions, features, and labels has to be chosen, while model_store has to be picked to store models. This is where Beaver is generic: you can use whatever model and data store you want, as long as they satisfy the necessary interfaces.
A strong design principle is that Beaver is technology agnostic. If you want to store data in Redis, or use MLFlow as a model store, then so be it. The only strong decision is that Beaver is written in Python.
Once instantiated, the Dam is supposed to be run on a server. There are then multiple ways to interact with it. The most common way would be to start an HTTP server, and interact with said server from a client. That is one way to use Beaver. But you can also directly use the Dam you instantiated, for instance if you don't need to have a client/server separation. This might be the case if you're deploying a model on an embedded device with no internet connection.
The easiest way is to run the provided docker-compose.yaml.
git clone https://github.com/MaxHalford/beaver
cd beaver
docker-compose upYou can also just install the beaver Python package.
pip install git+https://github.com/MaxHalford/beaverThen it's up to you to handle the rest. The first thing to do is initialize a Dam.
import beaver
dam = beaver.Dam(
model_store=beaver.model_store.ShelveModelStore('~Downloads'),
data_store=beaver.data_store.SQLDataStore('sqlite:///db.sqlite'),
training_regimes=[beaver.training.Regime.ASAP]
)
dam.build()The build method makes sure each component is ready to be used. Assuming the above code is in a file named server.py, you may then start an HTTP server by leveraging FastAPI:
uvicorn server:dam.http_server --port 3000The server can then be interacted with via an HTTP client:
import beaver
from river import linear_model
model = linear_model.LogisticRegression()
client = beaver.HTTPClient(host='http://127.0.0.1:3000')
client.models.upload('my_model', model)docker build -t beaver .
docker run -it -p 3000:3000 beaverComing soon.
git clone https://github.com/online-ml/beaver
cd beaver
pip install poetry
poetry install
poetry shell
pytestBeaver is free and open-source software licensed under the 3-clause BSD license.