Skip to content
This repository has been archived by the owner on Sep 13, 2023. It is now read-only.
/ mlem Public archive

๐Ÿถ A tool to package, serve, and deploy any ML model on any platform. Archived to be resurrected one day๐Ÿคž

License

Notifications You must be signed in to change notification settings

iterative/mlem

Folders and files

NameName
Last commit message
Last commit date

Latest commit

ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 

Repository files navigation

image

Check, test and release codecov PyPi License: Apache 2.0

MLEM helps you package and deploy machine learning models. It saves ML models in a standard format that can be used in a variety of production scenarios such as real-time REST serving or batch processing.

  • Run your ML models anywhere: Wrap models as a Python package or Docker Image, or deploy them to Heroku, SageMaker or Kubernetes (more platforms coming soon). Switch between platforms transparently, with a single command.

  • Model metadata into YAML automatically: Automatically include Python requirements and input data needs into a human-readable, deployment-ready format. Use the same metafile on any ML framework.

  • Stick to your training workflow: MLEM doesn't ask you to rewrite model training code. Add just two lines around your Python code: one to import the library and one to save the model.

  • Developer-first experience: Use the CLI when you feel like DevOps, or the API if you feel like a developer.

Why is MLEM special?

The main reason to use MLEM instead of other tools is to adopt a GitOps approach to manage model lifecycles.

  • Git as a single source of truth: MLEM writes model metadata to a plain text file that can be versioned in Git along with code. This enables GitFlow and other software engineering best practices.

  • Unify model and software deployment: Release models using the same processes used for software updates (branching, pull requests, etc.).

  • Reuse existing Git infrastructure: Use familiar hosting like Github or Gitlab for model management, instead of having separate services.

  • UNIX philosophy: MLEM is a modular tool that solves one problem very well. It integrates well into a larger toolset from Iterative.ai, such as DVC and CML.

Usage

This a quick walkthrough showcasing deployment functionality of MLEM.

Please read Get Started guide for a full version.

Installation

MLEM requires Python 3.

$ python -m pip install mlem

To install the pre-release version:

$ python -m pip install git+https://github.com/iterative/mlem

Saving the model

# train.py
from mlem.api import save
from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import load_iris

def main():
    data, y = load_iris(return_X_y=True, as_frame=True)
    rf = RandomForestClassifier(
        n_jobs=2,
        random_state=42,
    )
    rf.fit(data, y)

    save(
        rf,
        "models/rf",
        sample_data=data,
    )

if __name__ == "__main__":
    main()

Codification

Check out what we have:

$ ls models/
rf
rf.mlem
$ cat rf.mlem
Click to show `cat` output
artifacts:
  data:
    hash: ea4f1bf769414fdacc2075ef9de73be5
    size: 163651
    uri: rf
model_type:
  methods:
    predict:
      args:
      - name: data
        type_:
          columns:
          - sepal length (cm)
          - sepal width (cm)
          - petal length (cm)
          - petal width (cm)
          dtypes:
          - float64
          - float64
          - float64
          - float64
          index_cols: []
          type: dataframe
      name: predict
      returns:
        dtype: int64
        shape:
        - null
        type: ndarray
    predict_proba:
      args:
      - name: data
        type_:
          columns:
          - sepal length (cm)
          - sepal width (cm)
          - petal length (cm)
          - petal width (cm)
          dtypes:
          - float64
          - float64
          - float64
          - float64
          index_cols: []
          type: dataframe
      name: predict_proba
      returns:
        dtype: float64
        shape:
        - null
        - 3
        type: ndarray
  type: sklearn
object_type: model
requirements:
- module: sklearn
  version: 1.0.2
- module: pandas
  version: 1.4.1
- module: numpy
  version: 1.22.3

Deploying the model

If you want to follow this Quick Start, you'll need to sign up on https://heroku.com, create an API_KEY and populate HEROKU_API_KEY env var (or run heroku login in command line). Besides, you'll need to run heroku container:login. This will log you in to Heroku container registry.

Now we can deploy the model with mlem deploy (you need to use different app_name, since it's going to be published on https://herokuapp.com):

$ mlem deployment run heroku app.mlem \
  --model models/rf \
  --app_name example-mlem-get-started-app
โณ๏ธ Loading model from models/rf.mlem
โณ๏ธ Loading deployment from app.mlem
๐Ÿ›  Creating docker image for heroku
  ๐Ÿ›  Building MLEM wheel file...
  ๐Ÿ’ผ Adding model files...
  ๐Ÿ›  Generating dockerfile...
  ๐Ÿ’ผ Adding sources...
  ๐Ÿ’ผ Generating requirements file...
  ๐Ÿ›  Building docker image registry.heroku.com/example-mlem-get-started-app/web...
  โœ…  Built docker image registry.heroku.com/example-mlem-get-started-app/web
  ๐Ÿ”ผ Pushing image registry.heroku.com/example-mlem-get-started-app/web to registry.heroku.com
  โœ…  Pushed image registry.heroku.com/example-mlem-get-started-app/web to registry.heroku.com
๐Ÿ›  Releasing app example-mlem-get-started-app formation
โœ…  Service example-mlem-get-started-app is up. You can check it out at https://example-mlem-get-started-app.herokuapp.com/

Contributing

Contributions are welcome! Please see our Contributing Guide for more details.

Thanks to all our contributors!

Copyright

This project is distributed under the Apache license version 2.0 (see the LICENSE file in the project root).

By submitting a pull request to this project, you agree to license your contribution under the Apache license version 2.0 to this project.