v0.8.0

OpenSoundscape 0.8.0 Release

This release represent a substantial update to opensoundscape, containing new features, bug fixes, and breaking changes to some parts of the API.

New Features:

The notebook New in OpenSoundscape 0.8.0 summarizes many of the new features, with code examples. New features include:

• logging progress of model training and prediction to Weights and Biases
• writing mp3 files (thanks to soundfile 1.1.0)
• training CNNs on long audio files without splitting into clips
• updated Audio metadata reading and writing
• Audio classmethods noise and silence
• Audio method normalize
• top level imports of most used classes. Eg, from opensoundscape import Audio, Spectrogram, CNN
• debug preprocessing by asking train() and predict() by passing raise_errors=True to raise instead of catch Preprocessing errors
• option to keep all intermediate versions of preprocessed sample with preprocessor.forward(...,trace=True)
• new augmentations in actions module: audio_random_gain and audio_add_noise
• added support for EfficientNet architecture
• new Audio object properties for .rms and .dBFS

Breaking changes:

• Audio.duration is now a property (use .duration not .duration()), as are several Spectrogram attributes that were previously methods
• CNN.predict() now only returns one argument, the dataframe of scores. Optionally, return a second argument with invalid samples list. Use predict_multi_target_labels or predict_single_target_labels from the metrics module to generate 0/1 outputs from continuous scores.
• a few arguments to functions have changed names and order

Merged Pull Requests

• update docs to clarify loading old models by @sammlapp in #572
• update docs to clarify loading old models by @sammlapp in #571
• Handles file with zero annotations by @jatinkhilnani in #593
• Updated BoxedAnnotations docstring by @jatinkhilnani in #594
• Minor documentation updates by @sammlapp in #608
• resolves #591 by @sammlapp in #612
• Adds Python 3.9 version to CI and a macos runner. by @louisfh in #595
• #552 Standardize references to 'model' object by @louisfh in #587
• fixes #558 with named args for Spectrogram methods by @louisfh in #585
• Support writing mp3 (soundfile 0.11) by @sammlapp in #613
• resolves #614 - avoid bug in load channels of mono by @sammlapp in #616
• Add WandB logging and Support training on a dataframe of clip labels by @sammlapp in #609
• log top-scoring samples to wandb (resolves #575) by @sammlapp in #619
• Feat 478 efficientnet and refactor cnn_architectures by @sammlapp in #620
• Bump dependencies by @sammlapp in #621
• Feat 569 audio by @sammlapp in #622
• Resolves #563 mps logit not implemented by @sammlapp in #626
• Add mix and concat by @sammlapp in #623
• explicit arguments for scipy.spectrogram by @sammlapp in #627
• use explicit arguments for softmax() by @sammlapp in #629
• add json comment-string metadata support by @sammlapp in #632
• Issue561 - Flat Module Structure in Docs by @syunkova in #638
• add audio_random_gain in actions by @sammlapp in #637
• modify predict to only return score df by @sammlapp in #628
• Audio add noise augmentation by @sammlapp in #639
• Issue 617 clipdf by @sammlapp in #631
• Issue 636 one output node for single-class classification by @sammlapp in #641
• Change default args from lists to tuple by @sammlapp in #644
• Issue543_v2: Use_pretrained is deprecated by @syunkova in #645
• Fixed - preprocess - actions by @jatinkhilnani in #570
• Issue 562 preprocessing raise errors by @louisfh in #646
• Issue_510_annotations by @sammlapp in #647
• Debug/trace feature by @jatinkhilnani in #596
• version 0.8.0 release by @sammlapp in #648

Full Changelog: v0.7.1...v0.8.0