Fedot

Hello, AutoML folk! We’re releasing a minor version of FEDOT that includes the following.

PyPi release: https://pypi.org/project/fedot/0.7.1/

Features & Enhancements:
- Update of [GOLEM](https://github.com/aimclub/GOLEM) (core framework) dependency to 0.3.1 version, that has some important features itself
- Added *Meta Rules* that automatically select best parameters based on dataset: with rules for early_stopping_generations, preset and cross-validation folds were added (1057)
- Improved API parameters and made documentation more clear and structured (1067, 1041)
- Improve test suite and its performance (1098)
- Improved DataMerger for textual data – now multi-column text table can be used with FEDOT (1052)

Bugfixes:
- Bug with wrong combinations of operations in pipelines for time series forecasting was fixed.
- Multiple initial assumptions support was fixed (1070)
- Various minor fixes

Hi, folk!

This release marks major change! Our team separated all the core modules (graph, adapter, optimizer, tuner etc.) into the separate project GOLEM (https://github.com/aimclub/GOLEM).

FEDOT now contains modules related to Data handling, preprocessing, machine learning logic like Pipeline (implementation of ML Graph), ML operations, ML metrics.

There're also few other changes:

- 1029 -- CLSTM model refactoring
- 1046 -- Transition of GitHub tests from Python 3.7 to 3.8

Hi all!

Importantly, release 0.6.2 marks the last self-sufficient version of FEDOT before transition to GOLEM optimization core (https://github.com/aimclub/GOLEM).

This release introduces a number of API enhancements and several bug fixes.
Enhancements:

- 1017 -- Now it's possible to enable memory analytic for debugging the framework performance.
- 1019 -- It became easier and more effective to work with large datasets.
- 1021 -- PipelineBuilder now can be used for any Graphs with a help of appropriate Adapter.
- 990 -- A number of user API improvements.
- 1025 -- Data preprocessing is now optional and can be disabled through new API parameter.
- 1031 -- Refactoring of Pipeline Node classes, that now are much simplified.

Bug fixes:

- 1010 -- Tuning is more benefitial now with correct metric deviation computation.
- 1012
- 1022
- 1023

Hi, folk!
We're making a new minor release with a number of improvements. This is an important release in a sense that this is a last release of self-contained FEDOT. The next major release will mark a separation of the optimizer core into the separate project.

New features, better quality & changes in API

- More intuitive predict interface for time series forecasting (930)
- Pipeline save/load now have more intuitive behavior (971)
- Early stopping criteria now can take timeout into considerations, and not only number of iterations (early_stopping_timeout api parameter)
- Graph nodes now can be accessed by name or uid (982)
- Tuner speed is better due to better initial params in the search space (985)

Enhancements and fixes:

- Fix inplace modification of data during data definition (resolves 943)
- Fix regression preprocessing (955)
- Less evaluation errors during population selection in corner cases (956)
- Fix getting suitable operations for multi ts (981)
- Integration tests are fixed & passing now
- More minor fixes & minor class interface refactorings
- Important fix for multi-objective optimization (996)

Documentation is extended

Architectural refactorings are continued:

- Better PipelineAdapter (941)
- Abstracting optimiser core (most tasks in issue 713 are done)
Notably, Serializer subsystem is now extendable (969)

Hi everyone!
We released a new major version of FEDOT - 0.6.0

It includes a lot of major changes:

- Improvement of API for multi-modal datasets and models;
- New `PipelineBuilder` (597) – that simplifies manual construction of ML Pipelines;
- Joblib was embedded as a multiprocessing backend (843). Data exchange between processes minimized (926);
- Embedding stratify k fold strategy for cases with imbalance data;
- New visualization of graphs, pipelines and optimisation history;

Also, this release contains by a lot of architectural refactorings of the framework:

- New Graph Adapter subsystem (876);
- Merging two different implementation of evolutionary optimizer (parameter-free & usual) into one `EvoGraphOptimizer` (687)
- Architectural refactorings of the Graph hierarchy (750)
- Introduce notions of `Objective` & `Fitness` (654) – classes that substitutes simple `float` metric values & abstract single vs. multi-objective metrics
- Refactored parameter classes – for more intuitive segregation of different parameters controlling optimization process (852)
- Refactored `DataMerger` facility
- Refactoring of selection operator implementation (918)

Also, there are various bug-fixes related to ML operations, evolutionary operators & internal Graph operations.

The most important changes:

- Cache support for the cross-validation implemented;
- AutoML can be run without a time limit;
- Graph operators improved;
- Multi-task pipelines processing improved;
- Custom parameters support for external optimizer;
- Time series processing improved;
- Multimodal table processing improved;
- Lightweight docker prepared;
- Isolation Forest added as new operation
- Major and minor bugs are fixed.