Python machine learning backend for Moodle

=======================================================

This package is used by Moodle's mlbackend_python plugin.

Info about packaging a new version in _readme_moodle.md: https://github.com/moodlehq/moodle-mlbackend-python/blob/master/readme_moodle.md

Requirements

• Python 3.7

Setup

Setup a virtual enviornment and install the package locally

virtualenv -p /usr/local/opt/python@3.7/bin/python3 venv
pip install -r requirements.txt
python setup.py develop

Tests

Run the tests and generate a coverage report

pytest --cov=moodlemlbackend tests/

How Moodle Integrates and uses this library

The core of the code in the Moodle PHP backend is located here: https://github.com/moodle/moodle/tree/master/lib/mlbackend/python

Which has two possible configiruations:

1. Moodle talks to a backend REST server that hosts this library/models
2. Moodle calls a python process directly to get/generates models

See - https://github.com/moodle/moodle/blob/master/lib/mlbackend/python/classes/processor.php#L97

1. As a result as interactions occur and models are generated they need to be stored and this is currently done via the TensorFlow Saver (https://www.tensorflow.org/api_docs/python/tf/compat/v1/train/Saver) which is similar to Pickle.
2. If using the REST API, there is an included Flask service however it should be noted that these are synchronous calls so long running HTTP requests can/will timeout. It is recommended to use the direct python process call for large datasets.

Flow of calls

1. REST or Python process call to training, at the conclusion the model is "saved"
2. REST or Python process call to predict to get actual predictions
3. [Optional] REST or Python call to evaluation to get models scores

Python Process Calls

When the python process is called directly the moduls read directly from the system arguments for their inputs

Flask REST Calls

1. All REST calls supply a Basic Authorization header which the Flask apps looks in the MOODLE_MLBACKEND_PYTHON_USERS env vars for logins to use
2. Flask app can be configured to be backed by local storage or Amazon S3
3. Flask app REST calls can then have data pulled from either source, implies data is saved there by the caller before this app being called

Notes

• Python 3.8 supports makes the TensorFlow library only support up tp version 2.2 and the library currently has version 1.5 only calls. We get around this by doing

import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()