mlserve

mlserve turns your python models into RESTful API, serves web page with form generated to match your input data.

It may be useful if one wants to demonstrate created predictive model and quickly integrate into existing application. Additionally UI is provided for input data (based on training dataframe) and simple dashboard.

Project is not complete but already usable, so no any guaranties on API or UI backward compatibility.

Online Demo

Several models deployed online using heroku.com/free free dynos. Free apps sleep automatically after 30 mins of inactivity so first request may take some time.

• https://young-ridge-56019.herokuapp.com

Full source code and instructions available here: https://github.com/ml-libs/mlserve-demo

Ideas

mlsserve is small using following design based on following ideas:

• Simplicity and ease of use is primary objective.
• Application consists of two processes: IO process that runs HTTP server and responsible for fetching and sending data, as well as serve UI, other process (worker) is doing CPU intensive work related to predictions calculations.

Features

• Model predictions serving via RESTful API endpoint.
• Model predictions serving via generated UI.
• Web page to simplify models usage.
• Automatic UI generation to match your input data.
• Simple dashboard for monitoring purposes.

Installation

Installation process is simple, just:

$ pip install git+https://github.com/ml-libs/mlserve.git

Example

To deploy model just follow following simple steps:

Save your model into pickle file:

with open('boston_gbr.pkl', 'wb') as f:
    pickle.dump(clf, f)

Use build_schema function to build UI representation of pandas dataframe, and save it as json file file:

import mlserve

data_schema = mlserve.build_schema(df)
with open('boston.json', 'wb') as f:
    json.dump(data_schema, f)

Create configuration file with following format:

models:
  - name: "boston_regressor"  # url friendly name
    description: "Boston GBR"  # optional model description
    model_path: "boston_gbr.pkl"  # path to your saved model
    data_schema_path: "boston.json"  # path to data representation
    target: "target"  # name of the target column

Serve model:

$ mlserve -c models.yaml

Thats it, model is available throw REST API, you can test is with curl command:

$ curl --header "Content-Type: application/json" --request POST
--data '[{"feature1": 1, "feature2": 2}]'
http://127.0.0.1:9000/api/v1/models/boston_gradient_boosting_regressor/predict

UI is available via http://127.0.0.1:9000

Supported Frameworks

• Scikit-Learn
• Keras (planning)
• PyTorch (planning)

Requirements

• Python 3.6+
• aiohttp