Skip to main content

Class that wraps the notion of AI model to include an automatic persistence of trained models, with a complete history of their trainings and the associated metrics

Project description

Public Project AllOnIAModel

The main purpose of this module is to provide data scientists with the :obj:~alloniamodel.model.AllOnIAModel class, that wraps the notion of AI model to include an automatic persistence of trained models, with a complete history of their trainings and the associated metrics. This allows for an easier monitoring of the model performances, and the prediction pipeline is lighter as it does not require to redefine every intermediary functions, like feature engineering.

This persistence is done through the :obj:~alloniamodel.model.AllOnIAModel.save method that pickles an instance of the class to S3 using cloudpickle <https://github.com/cloudpipe/cloudpickle>_. Then, creating a model with a name that already exists on S3 will load it automatically.

This allows for an online monitoring of the model metrics, learning after learning, prediction after prediction.

The user can trigger training and prediction pipelines through the :obj:~alloniamodel.model.AllOnIAModel.learn and :obj:~alloniamodel.model.AllOnIAModel.apply methods (see :ref:pipeline_steps).

Most methods of the training or prediction pipelines will accept custom keyword arguments, which allows :obj:~alloniamodel.model.AllOnIAModel to cover a wide range of use-cases. See :ref:custom_keyword.

Even though it is a public package, it is not intended to be used outside AllOnIA's plateform.

You can find the user documentation at this URL

This is a public project. Everyone is welcome to contribute to it.

Basic usage

To use this class, the user needs to provide some mandatory inputs :

Mandatory user-defined instance attributes for learning

Assuming an instance of the class was created like that :

from alloniamodel import AllOnIAModel

model = AllOnIAModel("iris_classification")

The following attributes/methods must be defined/called for the instance :

  • :obj:~alloniamodel.model.AllOnIAModel.set_variables
predictives = ("feature 1", "feature 2", "feature 3", "feature 4")
targets = ("target 1",)
model.set_variables(predictives, targets)

Note that there is a default value for predictive and target variables, ("x",) and ("y",), but it will most of the time only be useful if using :obj:~alloniamodel.utils.SpecialDataFormat objects.

  • :obj:~alloniamodel.model.AllOnIAModel.model or :obj:~alloniamodel.model.AllOnIAModel.model_class
model.model = KNeighborsClassifier(n_neighbors=1)
# OR
model.model_class = KNeighborsClassifier

The user can specify any kind of model here, as long as it is a class with the fit and predict methods. The name of the fit and predict methods are respectively fit and predict by default but can be changed through the :obj:~alloniamodel.model.AllOnIAModel.fit_function_name and :obj:~alloniamodel.model.AllOnIAModel.predict_function_name attributes. The fit method should accept X as first argument, and y as second, if target variables were specified (that is why :obj:~alloniamodel.model.AllOnIAModel.set_variables must be called before setting the model). If no target variables were specified, it is assumed that the given model does not accept a y argument in its fit method. The predict method should accept X as first argument.

  • :obj:~alloniamodel.model.AllOnIAModel.raw_set (see :ref:data_input)

Once those are defined, the user can do

model.learn()
model.save()

But the user might also want to specify more things :

Optional user-defined instance attributes for learning

  • :obj:~alloniamodel.model.AllOnIAModel.add_validator
model.add_validator("surname name", "mail@adress", "admin")

Validators are not implemented yet, but in a futur update, any training or prediction will trigger a reporting sent to the specified adresses.

  • :obj:~alloniamodel.model.AllOnIAModel.train_val_test_split_function (see :ref:split)
from mytools import some_split_function

model.train_val_test_split_function = some_split_function
  • :obj:~alloniamodel.model.AllOnIAModel.set_set_sizes
# Set the validation and test set sizes, as fraction of the raw set size.
model.set_set_sizes(0.1, 0.2)
  • :obj:~alloniamodel.model.AllOnIAModel.feature_engineering_function (see :ref:feature_engineering)
from mytools import some_feature_engineering_function

model.feature_engineering_function = some_feature_engineering_function
  • :obj:~alloniamodel.model.AllOnIAModel.compute_metrics_function (see :ref:evaluating)
from mytools import some_compute_metrics_function

model.compute_metrics_function = some_compute_metrics_function

Mandatory user-defined instance attributes for prediction

  • :obj:~alloniamodel.model.AllOnIAModel.observations_set (see :ref:data_input)

Optional user-defined instance attributes for prediction

  • :obj:~alloniamodel.model.AllOnIAModel.postprocess_function (see :ref:pipeline_predict)
from mytools import some_postprocess_function

model.postprocess_function = some_postprocess_function

Simple learning example

Here you can find detailed notebooks, custom functions and prediction pipelines examples : :ref:examples.

Monitoring

See :ref:monitoring.

Installation

pip install alloniamodel

Contributing

This is an open-source project. Everyone is welcome to contribute to it. To do so, fork the repository, add your features/fixes on your forked repository, then open a merge request to the original repository.

Install dependencies using poetry

This project uses Poetry to manage its working environment. Install it before coding in the project.

Then, run

poetry env use python3.12
poetry install
poetry run pre-commit install

Testing

Tests are separated into several groups, that can require different packages.

You can run them all using tox:

poetry run pytest

Coverage

We use pytest-cov to display the coverage, so, after run tests you can check the reports (term, html, xml are enabled), if you want to improve your coverage, the better thing to do is to check the html report in your browser:

open htmlcov/index.html

Lint

To run the linters used by this project, you can run:

poetry run pre-commit run # Run lint only on staged files

# Manually check conventional commits format:
poetry run pre-commit run gitlint --hook-stage commit-msg --commit-msg-filename .git/COMMIT_EDITMSG

User documentation

The documentation source files are located in here. If you add new features, please add them to the documentation as well.

You can buid the documentation locally by doing

cd docs
make html

The produced documentation should then be readable by opening the file in docs/build/html/index.html in a web browser.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

alloniamodel-1.4.5.tar.gz (51.8 kB view details)

Uploaded Source

Built Distribution

alloniamodel-1.4.5-py3-none-any.whl (51.9 kB view details)

Uploaded Python 3

File details

Details for the file alloniamodel-1.4.5.tar.gz.

File metadata

  • Download URL: alloniamodel-1.4.5.tar.gz
  • Upload date:
  • Size: 51.8 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: poetry/1.8.3 CPython/3.12.1 Linux/5.4.0-148-generic

File hashes

Hashes for alloniamodel-1.4.5.tar.gz
Algorithm Hash digest
SHA256 2103ba130e92308f51158412f6186d3c665e0930758ded06e0e36fc2d716f1ea
MD5 c9da6a15537e3c96ab1f4feaed9ed193
BLAKE2b-256 03439d390615cf68515eb8780edefbee97a5778c17dd0a0ea5f5f132e9063e72

See more details on using hashes here.

File details

Details for the file alloniamodel-1.4.5-py3-none-any.whl.

File metadata

  • Download URL: alloniamodel-1.4.5-py3-none-any.whl
  • Upload date:
  • Size: 51.9 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: poetry/1.8.3 CPython/3.12.1 Linux/5.4.0-148-generic

File hashes

Hashes for alloniamodel-1.4.5-py3-none-any.whl
Algorithm Hash digest
SHA256 63afcb62df4781dec2153ddb186f9ee8c052393c66f1b5d2b507902f8e585fc6
MD5 da99c0805a795f4eafa76cd41e499d2e
BLAKE2b-256 6eb1e9557fd4c4909860adf0d998beef915ee940f19fb9431d244251fc6da368

See more details on using hashes here.

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page