0.14.0

0.14.0 (2022-07-13)

• Standarized formatting with Black
• Split up ORM into a standalone swappable backend
• Persistables maintain weakrefs for lineage
• Persistables are normal python objects now
• Hashing flag to reject non-serializable objects

What's Changed

• Black formatting by @eyadgaran in #103
• dask tweaks by @eyadgaran in #104
• Orm separation by @eyadgaran in #99
• Allow user to request a raised exception if hash(content) will be inconsistent by @ptoman-pa in #107
• version bump and changelog by @eyadgaran in #109

New Contributors

• @ptoman-pa made their first contribution in #107

Full Changelog: 0.13.0...0.14.0

0.13.0

• Path existence check for pandas serialization

What's Changed

• Path creation by @eyadgaran in #101

Full Changelog: 0.12.0...0.13.0

0.12.0

• Changed internal dataset structure from mixins to direct inheritance
• Condensed all pandas dataset types into a single base class
• Adds support for dask datasets
• Placeholders for additional dataset libraries
• Adds hashing support for dask dataframes
• Refactored persistence ("save_patterns") package into standalone extensible framework
• Adds context manager support to registries for temporary overwrite
• Refactor pipelines into library based subclasses

BREAKING CHANGES

• Pandas dataset will default param squeeze_return to False (classes expecting to return a series will need to be updated)
• Numpy dataset is considered unstable and will be redesigned in a future release
• Onedrive, Hickle, and database save patterns are removed (functionality is still available but a composed pattern is not predefined. these can be trivially added in user code if needed)
• Changed pandas hash output to int from numpy.int64 (due to breaking change in NumpyHasher)
• Changed primitive deterministic hash from pickle to md5
• Extracted data iterators into utility wrappers. Pipelines no longer have flags to return iterators
• Random split defaults are computed at runtime instead of precalculated (affects hash)

What's Changed

• Ml management structure by @eyadgaran in #87
• Python eol by @eyadgaran in #92
• Dataset libraries by @eyadgaran in #90
• Pipeline refactor by @eyadgaran in #96
• additional testing coverage by @eyadgaran in #83
• Adding Ensemble Model Histogram-based Gradient Boosting Classifier by @aolopez in #91
• version bump by @eyadgaran in #98

New Contributors

• @aolopez made their first contribution in #91

Full Changelog: 0.11.0...0.12.0

0.11.0

• Added support to hasher for initialized objects
• Adds support for arbitrary dataset splits and sections
• Dataset hooks to validate dataframe setting
• Pipelines no longer cache dataset splits and proxy directly to dataset on every call
• Introduces pipeline splits as reproducible projections over dataset splits
• Database utility to recalculate hashes for existing persistables

BREAKING CHANGES

• Hash for an uninitialized class changed from repr(cls) to "cls._module.cls._name"
• Database migrations no longer recalculate hashes. That has to be done manually via a utility

0.10.0

• Dataset external file setter with validation hooks
• Pandas changes to always return dataframe copies (does not extend to underlying python objects! eg lists, objects, etc)
• Pandas Dataset Subclasses for Single and Multi label datasets
• PersistableLoader methods do not require name as a parameter

BREAKING CHANGES

• PandasDataset is deprecated and will be dropped in a future release. Use SingleLabelPandasDataset or MultiLabelPandasDataset instead
• Pandas Dataset Classes require dataframe objects of type pd.DataFrame and will validate input (containers of pd.DataFrames are no longer supported)

0.9.3

• Patch release to support breaking changes in sqlalchemy 1.4

0.9.2

• minor patches

0.9.1

• Added cli with alembic support

0.9.0

• Refactored save patterns. Supports multiple concurrent save locations and arbitrary artifact declaration
• Registry centric model for easier extension and third party contrib
• Support for in-memory sqlite db
• Changed database JSON mapping class and dependency to support mutability tracking
• New import wrapper class to manage optional dependencies
• Added dataset_id as a Metric reference. Breaking workflow change! Will raise an error if a dataset is not added and the metric depends on it
• Dropped default Train pipeline split. Will return an empty split for split pipelines and a singleton full dataset split for NoSplitPipelines
• Explicitly migrated to tensorflow 2 and tf.keras

0.8.1

• Minor patches for classification metrics