Bringing Machine Learning models near your data, not the other way around!

Datasette ML is a Datasette plugin providing an MLOps platform to train, evaluate and make predictions from machine learning models.

All the underlying features are provided by sqlite-ml.

WARNING: this plugin is still experimental and not ready for production. Some breaking changes might happen between releases before reaching a stable version. Use it at your own risks!

Install this plugin in the same environment as Datasette:

$ datasette install datasette-ml

Define configuration within metadata.yml / metadata.json:

plugins:
  datasette-ml:
    db: sqml

A new menu entry is now available, pointing at /-/ml to access the MLOps dashboard.

Using datasette-ml you can start training Machine Learning models directly along your data, simply by using custom SQL functions! Let's get started by training a classifier against the famous "Iris Dataset" to predict flower types.

First let's load our data. For a real world project, your data may live with its own table or being accessed through an SQL view. For the purpose of this tutorial, we can use the sqml_load_dataset function to load standard Scikit-Learn datasets:

SELECT sqml_load_dataset('iris') AS dataset;

It will return the following data:

The Iris dataset is loaded into a table nammed dataset_iris, containing 150 examples, 4 features and 3 classes to be predicted.

Now that our dataset is ready, let's train a first machine learning model to perform a classification task using the sqml_train function:

SELECT sqml_train(
  'Iris prediction',
  'classification',
  'logistic_regression',
  'dataset_iris',
  'target'
) AS training;

It will return the following data:

We have just trained our first machine learning model! The output data informs us that our model has been trained, yields a score of 0.94 and has been deployed.

Now that we have trained our classifier, let's use it to make predictions!

Predict the target label for the first row of dataset_iris using the sqml_predict function:

SELECT
  dataset_iris.*,
  sqml_predict(
    'Iris prediction',
    json_object(
      'sepal length (cm)', [sepal length (cm)],
      'sepal width (cm)', [sepal width (cm)],
      'petal length (cm)', [petal length (cm)],
      'petal width (cm)', [petal width (cm)]
    )
  ) AS prediction
FROM dataset_iris
LIMIT 1;

This will output the following data:

Yay! Our prediction is matching the target label!

Let's see if we can find some predictions not matching the target label. To perform lots of predictions, we will use sqml_predict_batch which is more efficient than sqml_predict:

SELECT
  dataset_iris.*,
  batch.value AS prediction,
  dataset_iris.target = batch.value AS match
FROM
  dataset_iris
  JOIN json_each (
    (
      SELECT
        sqml_predict_batch(
          'Iris prediction',
          json_group_array(
            json_object(
              'sepal length (cm)', [sepal length (cm)],
              'sepal width (cm)', [sepal width (cm)],
              'petal length (cm)', [petal length (cm)],
              'petal width (cm)', [petal width (cm)]
            )
          )
        )
      FROM
        dataset_iris
    )
  ) batch ON (batch.rowid + 1) = dataset_iris.rowid
WHERE match = FALSE;

This will yield the following output data:

Oh no! 4 predictions have not predicted the correct target label!

Let's see if we can train a better algorithm to enhance the prediction quality.

Let's use a Support Vector Machine algorithm, usually yielding better results compared to the more simplistic Logistic Regression:

SELECT sqml_train(
  'Iris prediction',
  'classification',
  'svc',
  'dataset_iris',
  'target'
) AS training;

This will yield the following data:

We can already see that the score of this new model is higher than the previous one and it has been deployed.

Let's try our new classifier on the same dataset:

SELECT
  dataset_iris.*,
  batch.value AS prediction,
  dataset_iris.target = batch.value AS match
FROM
  dataset_iris
  JOIN json_each (
    (
      SELECT
        sqml_predict_batch(
          'Iris prediction',
          json_group_array(
            json_object(
              'sepal length (cm)', [sepal length (cm)],
              'sepal width (cm)', [sepal width (cm)],
              'petal length (cm)', [petal length (cm)],
              'petal width (cm)', [petal width (cm)]
            )
          )
        )
      FROM
        dataset_iris
    )
  ) batch ON (batch.rowid + 1) = dataset_iris.rowid
WHERE match = FALSE;

This will lead the following results:

Yay! We manage to predict one more target label with this new model!

Also note that we did not have to do anything to switch to the better model: exactly the same query is used to perform the prediction without having to specify anything about the new model! This is because new models are deployed automatically for the current experiment only if their score outperforms the score of the previously deployed model.

This plugin registers a few SQL functions to perform machine learning model training and predictions:

sqml_load_dataset(name, table)

• name: str: name of the dataset to load
• table: str: (optional) custom table name destination for the dataset

sqml_train(experiment_name, prediction_type, algorithm, dataset, target, test_size, split_strategy):

• experiment_name: str: name of the experiment to train the model within
• prediction_type: str: prediction task type to be performed for this experiment (regression, classification)
• algorithm: str: algorithm type to be trained
• dataset: str: name of the table or view containing the dataset
• target: str: name of the column to be treated as target label
• test_size: float: (optional) dataset test size ratio (default is 0.25)
• split_strategy: str: (optional) dataset train/test split strategy (default is shuffle)

sqml_predict(experiment_name, features)

• experiment_name: str: name of the experiment to train the model within
• features: json object: JSON object containing the features

sqml_predict_batch(experiment_name, features)

• experiment_name: str: name of the experiment to train the model within
• features: json list: JSON list containing all feature objects

To set up this plugin locally, first checkout the code. Then create a new virtual environment and the required dependencies:

poetry shell
poetry install

To run the QA suite:

black --check datasette_ml tests
flake8 datasette_ml tests
mypy datasette_ml tests
pytest -v --cov=datasette_ml --cov=tests --cov-branch --cov-report=term-missing tests

With the developmnent environment setup, you can run the demo locally:

python demo/generate.py
datasette --metadata demo/metadata.yml demo/sqml.db

All the things on the internet that have been inspiring this project:

• PostgresML
• MLFlow
• SQLite Run-Time Loadable Extensions
• Alex Garcia's sqlite-loadable-rs
• Alex Garcia's SQLite extensions
• Alex Garcia, "Making SQLite extensions pip install-able"
• Max Halford, "Online gradient descent written in SQL"
• Ricardo Anderegg, "Extending SQLite with Rust"

Licensed under Apache License, Version 2.0

Copyright (c) 2023 - present Romain Clement