Fedot v.0.3.1 has been released!

[J3FALL]

Hi everyone!
During the last month, we have merged several major features and fixed a bunch of bugs. Some of them are experimental and should be tested extensively in real-world cases. But we have tried our best and covered it with unit tests.

The new version is available and can be imported via pip: pip install fedot==0.3.1
The most important changes:

• ML pipelines for multi-modal datasets

Usually, you have a heterogeneous dataset with various types, such as images, numerical properties, texts, etc. And you want to build a ML pipeline for predictions using all data you have. We have implemented a prototype of such pipelines in FEDOT.
AutoML for this purpose should work too if you build a Multi-Modal pipeline as an initial solution of Composer and run the optimization. The example is available here.
Currently, it is supported only via extended initialization, but it would be available through FEDOT API soon.

• Decompose operation in ML pipelines

In some cases, during the modeling researchers are trying to take into account the error between the observations and model predictions. Usually, they are even to model this error. A similar idea is implemented in boosting method.
In the release, we have implemented this feature, see this PR for details.
Also, some examples are available that show how to use this method in classification, regression, and time series forecasting tasks. We did not test the decompose feature in Composer, but it could work if this operation is given in secondary available operations. A nice opportunity to test it and give feedback through issues!

• Cross-validation in Composer

Later in Composer the dataset is split simply on train/test samples. Now, K-Fold CV can be used as an optimized metric. The overall quality of the final pipeline could be better, but the performance is obviously lower.
In order to use it in API, just set cv_folds = True in Fedot initialization.

• Add Memory and Time profilers

We have found out that for debugging of AutoML, profiling could show the pitfalls of the code. After weeks of discovery of profiling tools, we have implemented two wrappers of profilers both for time and memory resources. Now, you can profile your computations using only two lines of code. See the examples here.

• Memory consumption has been improved

We also found out that during the composing, a memory leak occurs. However, if we are trying to release the memory of Chain objects that are not actual already during the evolution, it improves the memory usage dramatically. For details, see the merged PR.

As always, thanks for your attention and stars, new issues and feedback are appreciated!
Cya!