Changelog¶

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog.

0.3.0 - 2020-02-23¶

Add support for shared steps (PR #19). Now steps can be called several times on different inputs.
- This is a backwards-incompatible change. The outputs of the steps now follow the following format: step_name:port/output_number. (Previously it was step_name/output_number)
Add option to include targets in plot_model (PR #20).
Add new fit_compute_func argument to Step.__call__ that allows to specify custom behavior at fit time (PR #22).
Add documentation built with Sphinx and hosted on baikal.readthedocs.io (PR #29).

Move compute_func (previously function) and trainable args to Step.__call__ (PR #18).
- Also, the default value is changed from None to "auto".
- This is a backwards-incompatible change.
Raise RuntimeError chained with the original exception in Model.fit and Model.predict.

Add clarification in that steps must be named in build_fn when using SKLearnWrapper
Fix bug where the compute function was not being transferred when replacing a step in Model.set_params.
Fix an API inconsistency regarding the handling of the arguments of fit/compute for steps with multiple inputs and targets (PR #21).
Fix several bugs in plot_model (it was largely broken) (PR #20, PR #24).

This CHANGELOG file.
Introduced new targets API (PR #1).
- Steps now take an optional targets argument at call time to specify inputs for target data at fit time.
- Correspondingly, Model also takes an additional argument for these targets.
- The extra_targets argument in Model.fit was removed.
Step enhancements
- make_step factory function to ease definition of steps.
- Added support for function arguments to Lambda step (PR #8).
- Added new Split step (PR #9).