v0.6.2

This release provides minor updates and bug fixes in the OpenSoundscape code base and documentation but does not make breaking changes to the API.

Merged Pull Requests:

• fix issue 438 cuda by @sammlapp in #454
• Issue 447 pandas dependency by @louisfh in #461
• Resolves Issue #457 parse metadata by @sammlapp in #464
• modify make_clip_df to handle a single filepath string nicely by @louisfh in #463
• Docs 0.6.2 issues #420 #433 #439 by @louisfh in #465
• Docs for 0.6.2: issues #443 #445 #462 by @sammlapp in #466
• Issue #453 lstsq by @sammlapp in #470
• v0.6.2 by @sammlapp in #471

Full Changelog: v0.6.1...v0.6.2

v0.6.1

This release adds new functionality to OpenSoundscape and resolves a few issues.

Highlights

• Saving with model.save() and loading with cnn.load_model() or cnn.load_outdated_model(): Models are now saved and loaded as the entire object, so that the loaded model object has all attributes of the object when it was saved. Models saved with previous versions of opensoundscape must be loaded with opensoundscape.torch.models.cnn.load_outdated_model - see the tutorial "Predicting with pretrained CNNs" for details
• Directly load segments of audio files: using the new arguments offset and duration of Audio.from_file() you can quickly load an arbitrary section of a .WAV file without loading the whole file. (It works for mp3s too but is not as fast in that case.)
• Directly predict on "splits" of files without actually splitting them: New functionality including the ClipLoadingSpectrogramPreprocessor class allows you to quickly and easily generate CNN predictions on short (eg 5 second) segments of long audio files without needing to split the files into clips first. See "Predicting with pretrained CNNs" tutorial for an example
• Github homepage gives quick-start examples for #tldr seekers
• model.device now allows you to move to a specific cuda device other than 'cuda:0'
• PytorchModel init allows user to specify name of architecture instead of architecture object, eg PytorchModel('resnet50',classes=[0,1])
• All models and architectures can be loaded without downloading pretrained weights. Resnet18Multiclass and Resnet18Binary have been refactored as subclasses of PytorchModel.

What's Changed

• add from_checkpoint() method to Resnet18Multiclass by @sammlapp in #413
• Add more informative Raven annotation file errors by @rhine3 in #419
• Issue 410 cwtpeaks by @sammlapp in #416
• Handle user-specified devices like 'cuda:1' by @sammlapp in #422
• Issue #411 randominit by @sammlapp in #424
• Feat fast load by @sammlapp in #423
• Make split_and_save a method of Audio object by @louisfh in #426
• Issue 427 load old models by @sammlapp in #428
• Allow models to be created using architecture name (string) by @sammlapp in #430
• update ML tutorials with new best practices and methods by @sammlapp in #431
• update init files to allow top-level import of opensoundscape by @sammlapp in #434
• add offset/duration example for Audio.from_file() by @sammlapp in #435
• Bug fixes, docs updates, and new prediction tools for 0.6.1 by @sammlapp in #432

Full Changelog: v0.6.0...v0.6.1

v0.6.0

This release represents a significant update to OpenSoundscape and includes changes to the API. It introduces new features and fixes various outstanding issues and bugs.

new feature highlights

• annotations module (replaces raven module) with new tools for manipulating audio annotations, such as those created in Raven software
• signal module for signal processing, including tool for Ruffed Grouse drumming detection
• directly load and predict on long audio files with CNNs without splitting first (splits audio into shorter clips in memory - this eliminates the need to split and save short clips for every long audio file before prediction) - see example in "Predicting with pre-trained CNNs" tutorial
• Audio.from_file now loads metadata from audio files into Audio object's .metadata dictionary. For files recorded by AudioMoths, this can include a recording timestamp and ambient temperature measurement, among other metadata
• audiomoth module assists in parsing AudioMoth recordings' metadata (e.g., ambient temperature and recording time) and file names (to datetime objects)

What's Changed: Pull Requests from Develop branch

• Bring develop "up to date" with master by @rhine3 in #334
• Update contributor instructions by @jkitzes in #356
• Issue 351: model training continuation by @sammlapp in #358
• Misc preprocessing updates by @sammlapp in #338
• Issue 364 dont resample by @sammlapp in #367
• Doc fixes by @jkitzes in #374
• Feat reduced filesize by @louisfh in #377
• Add ruffed grouse detection method by @sammlapp in #380
• Add functionality to split audio while predicting with CNN by @sammlapp in #381
• Add function for ruffed grouse detection by @sammlapp in #384
• Issue 287 moth tools by @sammlapp in #390
• Resolve a few preprocessing/training bugs by @sammlapp in #394
• annotation module by @sammlapp in #386
• Issue 378 ribbit bug by @sammlapp in #389
• Load metadata with Audio.from_file by @sammlapp in #398
• Remove species_table.py module by @louisfh in #396
• Remove commands module by @louisfh in #402
• Update docs modules and fix bugs by @sammlapp in #406
• Issue 294 per-class thresholds for binary predictions by @louisfh in #399
• unpin dependencies and add pywavelets - all tests pass py=3.7 or 3.8 by @sammlapp in #408
• Version 0.6.0 by @sammlapp in #409

New Contributors

• @jkitzes made their first contribution in #356
• @louisfh made their first contribution in #377

Full Changelog: v0.5.0...v0.6.0

v0.5.0

This release represents a significant update to OpenSoundscape and a major change to the OpenSoundscape API. It includes the following features:

• Training with multi-class models
• Customizable CNN parameters:
  • learning rate and scheduling
  • regularization weight decay
  • pretrained weights
  • ability to freeze the feature extractor
• Customizable CNN architectures
  • pre-supplied classes for ResNet18 and Inceptionv3
  • General PytorchModel class that can be used with a variety of other provided architectures
• Customizable preprocessing pipelines
• Notebooks documenting the new API and customization

v0.4.7: Merge pull request #267 from kitzeslab/develop

Release v0.4.7