Changes¶

v0.1.51¶

Fixed bug tha threw a TypeError if multiple algorithms were run in a single JAMS file with None and other label_ids in it
Added new vmo oracle segmentation method (by Cheng-i Wang, thanks!)
Adapting sonify function to latest numpy
Using KMeans from sklearn instead of scipy for 2D-FMC. Results are better
Making sure we are never using more number of clusters than number of segments for 2D-FMC
Added new parameter 2dfmc_offset in the 2D-FMC method
Using np.inf normalization for 2D-FMC now, since it seems to yield better results (at least for Beatles TUT)
Padding beat-sync features now, seems to fix potential misalignment of boundaries. Some algorithms (2D-FMC, CNMF) seem to yield better results now
Modified features file: two new fields may be addeed: est_beatsync_times and ann_beatsync_times.
The member variable _framesync_times in the Features was never updated. Fixed it

Fixed bug of selecting framesync features
OLDA and Scluster hierarchies are consistent now (first element in hierarchy in the highest in both algorithms; this was also true for Scluster before this fix)
Warning message is displayed if jams file exists but can’t be read during features computation
If two algorithms used at the same time have the same name, and AssertionError is raised
Fixed normalization problem: now algorithms have independent normalization parameters
Adding out_file variable in main process function of the eval module
Reporting proper weighted F-measure for the perceptual Hit Rate
Fixed bug of annotator id not correctly passed to hierarchical evaluation function
More unit tests
Added script to upload to Pypi

Adapting code to librosa 0.5
Improved coveralls (starting testing plots)
Allowing computation of label algorithms for each layer of hierarchical boundaries
Removed old notebooks that were either useless or belonged to msaf-data
Removed old ds-name references, improving code health