Release 0.20

This technical release focuses on stability and model robustness improvements.

New features

• Custom metrics support
  Users can now define and register custom evaluation metrics within downstream models. This includes full compatibility with torchmetrics, duplicate-name checks, and consistent integration across training, validation, and monitoring phases.

---

Improvements

• Improved stability and speed of real-time inference
  Optimized the inference server interface and pipeline for more stable initialization, better resource utilization, and faster CI execution.

• Improved numerical stability of training on highly multimodal data
  Enhanced buffer sampling and shuffling to ensure better coverage of training examples, smoother convergence, and improved overall training stability.

• Improved regression and classification predictions
  Revised the random splitting strategy and application a uniform mixture to raw scores, leading to more balanced score distributions and reduced training bias.

---

Fixes

• Sketch width and optimal depth alignment
  Prevented potential crashes caused by conflicting sketch dimensions when handling certain class counts.

• Date parsing for time series not working with formatted data
  Fixed an issue where date columns were not parsed or sanitized in time-series data when a date format was provided.

• Invalid one-hot recommendation metrics for low number of candidates
  Prevented crashes on OneHotRecommendationTask models when the candidate pool was smaller than k.

• Text embedding stability
  Fixed occasional crashes during text model training under specific edge conditions.

• Short time-series handling
  Resolved a crash that occurred when the time-series length was shorter than the kernel size.

• Insufficient training data handling
  Introduced graceful exit with a clear and informative message when the available data is insufficient to fill a full batch across all selected devices.