ballet 0.19.5

Core functionality for lightweight, collaborative data science projects

ballet

A lightweight framework for collaborative, open-source data science projects through feature engineering.

• Free software: MIT license
• Documentation: https://ballet.github.io/ballet
• Homepage: https://github.com/ballet/ballet

Overview

Do you develop machine learning models? Do you work by yourself or on a team? Do you share notebooks or are you committing code to a shared repository? In contrast to successful, massively collaborative, open-source projects like the Linux kernel, the Rails framework, Firefox, GNU, or Tensorflow, most data science projects are developed by just a handful of people. But think if the open-source community could leverage its ingenuity and determination to collaboratively develop data science projects to predict the incidence of disease in a population, to predict whether vulnerable children will be evicted from their homes, or to predict whether learners will drop out of online courses.

Our vision is to make collaborative data science possible by making it more like open-source software development. Our approach is based on decomposing the data science process into modular patches that can then be intelligently combined, representing objects like "feature definition", "labeling function", or "prediction task definition". Collaborators work in parallel to write patches and submit them to a repo. The core Ballet framework provides the underlying functionality to merge high-quality contributions, collect modules from the file system, and compose the accepted contributions into a single product. It also provides Assemblé, a familiar notebook-based development experience that is friendly to data scientists and other inexperienced open-source contributors. We don't require any computing infrastructure beyond that which is commonly used in open-source software development.

Currently, Ballet focuses on supporting collaboratively developing feature engineering pipelines, an important part of many data science projects. Individual feature definitions are represented as separate Python modules, declaring the subset of a dataframe that they operate on and a scikit-learn-style learned transformer that extracts feature values from the raw data. Ballet collects individual feature definitions and composes them into a feature engineering pipeline. At any point, a project built on Ballet can be installed for end-to-end feature engineering on new data instances for the same problem. How do we ensure the feature engineering pipeline is always useful? Ballet thoroughly validates proposed feature definitions for correctness and machine learning performance, using an extensive test suite and a novel streaming feature definition selection algorithm. Accepted feature definitions can be automatically merged by the Ballet Bot into projects.

Next steps

Are you a data owner or project maintainer that wants to organize a collaboration?

👉 Check out the Ballet Maintainer Guide

Are you a data scientist or enthusiast that wants to join a collaboration?

👉 Check out the Ballet Contributor Guide

Want to learn about how Ballet enables Better Feature Engineering™️?

👉 Check out the Feature Engineering Guide

Want to see a demo collaboration in progress and maybe even participate yourself?

👉 Check out the ballet-predict-house-prices project

Source code organization

This is a quick overview to the Ballet core source code organization. For more information about contributing to Ballet core itself, see here.

path	description
cli.py	the ballet command line utility
client.py	the interactive client for users
contrib.py	collecting feature definitions from individual modules in source files in the file system
eng/base.py	abstractions for transformers used in feature definitions, such as BaseTransformer
eng/{misc,missing,ts}.py	custom transformers for missing data, time series problems, and more
eng/external.py	re-export of transformers from external libraries such as scikit-learn and feature_engine
feature.py	the Feature abstraction
pipeline.py	the FeatureEngineeringPipeline abstraction
project.py	the interface between a specific Ballet project and the core Ballet library, such as utilities to load project-specific information and the Project abstraction
templates/	cookiecutter templates for creating a new Ballet project or creating a new feature definition
templating.py	user-facing functionality on top of the templates
transformer.py	wrappers for transformers that make them play nicely together in a pipeline
update.py	functionality to update the project template from a new upstream release
util/	various utilities
validation/main.py	entry point for all validation routines
validation/base.py	abstractions used in validation such as the FeaturePerformanceEvaluator
validation/common.py	common functionality used in validation, such as the ability to collect relevant changes between a current environment and a reference environment (such as a pull request vs a default branch)
validation/entropy.py	statistical estimation routines used in feature definition selection algorithms, such as estimators for entropy, mutual information, and conditional mutual information
validation/feature_acceptance/	validation routines for feature acceptance
validation/feature_pruning/	validation routines for feature pruning
validation/feature_api/	validation routines for feature APIs
validation/project_structure/	validation routines for project structure

History

0.19.5 (2021-07-17)

• Fix bug with deepcopying ballet.pipeline.FeatureEngineeringPipeline

0.19.4 (2021-07-17)

• Fix bug with deepcopying ballet.eng.base.SubsetTransformer (#90)
• Add ballet.drop_missing_targets primitive

0.19.3 (2021-06-28)

• Support missing targets in discovery and feature performance evaluation (#89)
• Add ninputs to summary statistics in ballet.discovery.discover

0.19.2 (2021-06-21)

• Improve discrete column detection in the case of many repeated values
• Add ncontinuous and ndiscrete to summary statistics in ballet.discovery.discover

0.19.1 (2021-06-20)

• Defer computation of some expensive summary statistics in ballet.discovery.discover

0.19.0 (2021-06-16)

• Support callable as feature input (#88)

0.18.0 (2021-06-06)

• Added Consumer Guide
• Can use Ballet together with MLBlocks to engineer features and then use additional preprocessing and ML components (#86)
• Can wrap the extracted feature matrix in a data frame with named columns derived from feature.output or feature.name
• Implemented ballet.encoder.EncoderPipeline to (mostly) mirror ballet.pipeline.FeatureEngineeringPipeline
• Can specify the dataset used for fitting the pipeline in the engineer-features CLI via --train-dir path/to/train/dir

0.17.0 (2021-05-24)

• Support nested transformers, both with nested features and with input/transformer tuples wrapped with SubsetTransformers (#82)
• Allow Client.discover to skip summary statistics if development dataset cannot be loaded or if features produce errors

0.16.0 (2021-05-22)

• Add Client.discover functionality (#80)
• Switch the order of NullFiller parameters to more closely resemble fillna signature

0.15.2 (2021-05-14)

• Operate columnwise in VarianceThresholdAccepter, rather than computing the variance of the entire feature group.

0.15.1 (2021-05-12)

• Add debug logging for new accepters

0.15.0 (2021-05-12)

• Add VarianceThresholdAccepter, MutualInformationAccepter, and CompoundAccepter (#76)

0.14.0 (2021-05-11)

• Support using holdout data splits in validation (#75)
• Fix CLI program name in projects (#74)
• Fix bug with load_config usage in python REPL (#73)
• Reorganize external feature engineering primitives to ballet/eng/external/**.py. Imports like from ballet.eng.external import MyPrimitive are unaffected.

0.13.1 (2021-04-02)

• Fix upgrade check in ballet update-project-template to migrate away from deprecated PyPI XML-RPC API.

0.13.0 (2021-03-30)

• Fix links in project template

0.12.0 (2021-03-10)

• Automate creation of GitHub repository in quickstart

0.11.0 (2021-03-04)

• Allow validation to be run from topic branches locally

0.10.0 (2021-02-23)

• Add Project.version property

0.9.0 (2021-02-16)

• Add support for managed branching via ballet start-new-feature --branching (defaults to enabled)
• Remove confusing ballet.project.config attribute
• Implement ballet.project.load_config as a better alternative, and use this in the project template's load_data

0.8.2 (2021-02-16)

• Fix bug with str(t) or repr(t) for DelegatingRobustTransformer

0.8.1 (2021-02-16)

• Fix bug with str(t) or repr(t) for SimpleFunctionTransformer

0.8.0 (2021-02-02)

• Fix bug with detecting updates to Ballet due to PyPI API outage
• Fix some dependency conflicts
• Reference ballet-assemble in project template
• Bump feature_engine to 1.0

0.7.11 (2020-09-16)

• Reduce verbosity of conversion approach logging by moving some messages to TRACE level
• Implement "else" transformer for ConditionalTransformer
• Improve GFSSF iteration logging

0.7.10 (2020-09-08)

• Fix bug with different treatment of y_df and y; now, y_df is passed to the feature engineering pipeline, and y is passed to the feature validation routines as applicable.
• Switch back to using Gitter

0.7.9 (2020-08-15)

• Add give_advice feature for FeatureAPICheck and other checks to log message on how to fix failure
• Improve logging of GFSSFAccepter and GFSSFPruner
• Improve __str__ for DelegatingRobustTransformer and consequently consumers
• Change default log format to SIMPLE_LOG_FORMAT
• Various bug fixes and improvements

0.7.8 (2020-08-13)

• Add CanTransformNewRowsCheck to feature API checks

0.7.7 (2020-08-12)

• Support None as the transformer in a Feature, it will be automatically converted to an IdentityTransformer
• Implement ColumnSelector
• Update docs
• Various bug fixes and improvements

0.7.6 (2020-08-12)

• Re-export feature engineering primitives from various libraries
• Show type annotations in docs
• Update guides
• Various bug fixes and improvements

0.7.5 (2020-08-03)

• Make validator parameters configurable in ballet.yml file (e.g. λ_1 and λ_2 for GFSSF algorithms)
• Support dynaconf 3.x

0.7.4 (2020-07-22)

• Accept logger names, as well as logger instances, in ballet.util.log.enable
• Updated docs

0.7.3 (2020-07-21)

• Add load_data method with built-in caching to project API
• Fix bug in GFSSF accepter
• Always use encoded target during validation
• Various bug fixes and improvements

0.7.2 (2020-07-21)

• Add sample analysis notebook to project template
• Add binder url/badge to project template
• Fix bug with enabling logging with multiple loggers

0.7.1 (2020-07-20)

• Add client for easy interactive usage (ballet.b)
• Add binder setup to project template

0.7 (2020-07-17)

• Revamp project template: update project structure, create single API via FeatureEngineeringProject, use and add support for pyinvoke, revamp build into engineer_features, support repolockr bot
• Improve ballet.project.Project: can create by ascending from given path, can create from current working directory, can resolve arbitrary project symbol, exposes project's API
• Check for and notify of new release of ballet during project update (ballet update-project-template)
• Add ComputedValueTransformer to ballet.eng
• Move stacklog to separate project and install it
• Add validators that {never,always} accept submissions
• Add feature API checks to ensure that the feature can fit and transform a single row
• Add feature engineering guide to documentation and significantly expand contributor guide
• Add bot installation instructions to maintainer guide
• Add type annotations throughout
• Drop support for py35, add support for py38
• Deprecate modeling code
• Various bug fixes and improvements

0.6 (2019-11-12)

• Implement GFSSF validators and random validators
• Improve validators and allow validators to be configured in ballet.yml
• Improve project template
• Create ballet CLI
• Bug fixes and performance improvements

0.5 (2018-10-14)

• Add project template and ballet-quickstart command
• Add project structure checks and feature API checks
• Implement multi-stage validation routine driver

0.4 (2018-09-21)

• Implement Modeler for versatile modeling and evaluation
• Change project name

0.3 (2018-04-28)

• Implement PullRequestFeatureValidator
• Add util.travis, util.modutil, util.git util modules

0.2 (2018-04-11)

• Implement ArrayLikeEqualityTestingMixin
• Implement collect_contrib_features

0.1 (2018-04-08)

• First release on PyPI