Release 1.1.1

This release contains only one change:

• Set self.normalized_specs and self.unnormalized_specs in Predictor::get_overlapping_scores.

This allows callers of get_overlapping_scores to access the spectrograms used to calculate scores.

Release 1.1.0

This release contains the following improvements to support frame-level labels, which improve the ability of SED classifier heads to accurately identify the start and end points of sounds:

• A new --occlude option in the analyze-db command runs occlusion sensitivity analysis (OSA) as a first step in deriving frame-level labels.
• A new pickle-frame command, which generates frame-level labels from the OSA output.
• A new frame_label_pickle config parameter is used in training to process frame-level labels.

In addition, this release contains the following minor changes:

• A new --dims option in the search command.
• Do not try to get spectrograms if < 0.5 seconds of audio, since that tends to cause downstream problems.

Release 1.0.8

This release contains just one change:

• Add optional quantile parameter to Predictor constructor. If specified, use given quantile for assigning SED frame scores to fixed-length segments. Otherwise use max frame.

Release 1.0.7

This release contains one change:

• In predictor.py, change get_overlapping_scores to give the caller more control over spectrogram offsets.

Release 1.0.6

This release contains the following changes:

1. Fix a bug where the --power argument was ignored in the plot-db command.
2. Revise the offset creation logic in Predictor::get_overlapping_scores.

Release 1.0.5

This release includes only one change:

• Allow inference to process the final segment of audio even if it is shorter than a full spectrogram, as long as at least 1 second of audio remains. The previous logic stopped generating start times too early, skipping audio near the end of a recording that didn't fit a full spectrogram-sized window.

Release 1.0.4

This release includes two small changes:

1. The scalable_sed head is renamed to reduced_sed, but scalable_sed is kept for backwards compatibility for now.
2. The max_models config parameter is moved from the the misc to the infer section.

Release 1.0.3

This release fixed the following bug:

• The rpt-test command sometimes terminates unexpectedly if there is a very short recording

Release 1.0.2

Changes:

1. In the rpt-labels command, rename summary.csv to classes.csv, change output sorting, and add recordings.csv output.
2. Update the help test for the pickle-occurrence command.

Release 1.0.1

This release fixes the following bug:

• When use_spec_cache=true (it is false by default), an exception may be thrown if the audio is very slightly longer than a multiple of the segment length