The majority of this release comprises new capabilities for distributed training using the Horovod framework.
- Support for distributed ML training using the Horovod framework.
- Unity client updated to with checkbox to enable running Horovod models.
- This feature has been tested using PyTorch.
- Updated Dask and Kubernetes pod management supporting distributed (Horovod) modeling.
- Added integration tests to analyze standard, image, and NLP ops on a daily basis.
- Added Luigi tests for no-code ML covering K-Means and Spectral clustering algorithms.
Updates and Fixes¶
- Predict/Explain bug-fix updated to apply to multi-VDS training as well.
- Fixed the encode operation error for tabular classification.
- Fixed the image transformation operations so that the processing of corrupted images are prevented.
- Reduction in VDS loading time.
- Jupyter notebook model creation supports model assignment to projects. Jupyter notebook capabilities will be surfaced in the Unity client at a later point.
- Model library/framework federated string (i.e., replace “FATE” w/ “federated” for all applicable cases) updated on the front-end and the back-end.
- Federation setup is now dynamic and does not depend on the configuration file. Federates can be added or removed at runtime.
- Vaious REST endpoints updated for performance improvements (reduced response time).