- Prepare for repo ownership transfer
- Custom metadata columns are not filtered out
- Fixed Pandas FutureDeprecation warning on frame.append
- Operating point arrays are initialised with np.nan
- Ground Truth and Metadata tables drop extra columns
- Fixed Numpy type deprecation warnings
- input tables with invalid indexes no longer create issues (fix for issue #3: PSDSEval.add_operating_point needs index column explicitly right now.)
- ground_truth table and detection tables are now checked for invalid events (events with onset > offset or with overlapping events from the same class). An error is raised if invalid events are found.
- zero length events are now automatically omitted from ground_truth and detection tables before the evaluation.
- tests and examples are updated to fix the overlapping labels in DCASE ground truth example
- Allow user to insert information per operating point
- Retrieve operating points based on given criteria
- introduced a new plot function for displaying per-class PSD-ROC.
- added support for external axes to the plot_psd_roc function.
- added the duration_unit as class attribute. It is then returned by the PSDSEval.psds function in the PSDS namedtuple.
- added a jupyter notebook that shows how the PSDS can be used to extract performance insights for the DCASE2020 Task 4 baseline system. It also explains a few features of PSDS metric in more detail.
- fixed a typo in README.md
- Added function for computing macro F-score. It is defined as the average of the per class F-score metrics.
- Relaxed check of increasing monotonicity for eTPR (effective True Positive Rate).
This check is performed in the function that calculates the area under the PSD-ROC
curve and it is now disabled when the parameter
alpha_ct > 0. In this case such property can't be guaranteed. - The confusion matrix (
counts) was built with inverse axis which then caused the cross-trigger rate to be assigned to the wrong class.