Feature Requests for 1.0

[kahst]

DONE --> We can now export as Raven model. DONE --> Export as single tflite model. We need to change the analysis with a custom classifier too if we do that.

Export re-trained model as Raven model / single model file

[kahst]

DONE but not tested --> We use a custom loss function to account for negative labels

Support negative training samples for re-training.

• folder names start with "-"
• compute loss for negative samples (inverse loss)
• mask loss for all other classes
• combine positive and negative loss as weighted sum

[kahst]

Multi-label samples for re-training

• maybe move from folder names to csv file or allow both

[kahst]

DONE --> We now support crop modes 'center', 'first' and 'segments'

Deal with training samples that are longer than 3-seconds

• split into consecutive 3-second chunks
• allow overlap for splitting

[kahst]

Installer for Windows and MacOS

[kahst]

DONE --> we now use AUPRC

Provide meaningful metric for re-training

• current metric implementation cannot deal with "all zero" label vectors, which we use for background classes
• metric also needs to support negative samples once implemented
• maybe move from metric to loss only

[kahst]

Signed Windows and MacOS build

See #160

[kahst]

Display model version

See #146

[kahst]

DONE --> we read high_freq from recording

Adjust max frequency in Raven selection table to sampling rate of source recording

See #140

[kahst]

DONE

Support multi-file selection review in Raven by adding Begin.File to selection table

See #58

[kahst]

DONE

Add species.py to GUI

[kahst]

What license should a re-trained model have? Can we provide a permissive, commercial-use license for re-trained models?

[kahst]

Somehow, results in single file mode vs. multiple file mode differ when using a custom classifier

[max-mauermann]

DONE --> Saves embeddings and labels to npz file

• Reduce data-loading time when training with same data (maybe save embeddings in temporary folder, so they can be used again, or enable training with embeddings as inputs)

[kahst]

In train.py allow to use Raven selection tables to crop training samples based on annotations from table based on segments.py code.

• train data does not have to be sorted into folders
• folder with train data should contain audio files and one Ravel selection table (or other format we support) for each file
• we extract training samples based on annotations in file
  • we'll pad samples that are too short and split samples that are too long
• segmets.py should contain the code necessary

[kahst]

Serverless demo webpage to explore the analyzer performance for an example recording based on Gradio Lite.

[kahst]

Can we get rid of these warnings when saving a model?

WARNING:absl:Importing a function (__inference_BLOCK_3-1_ACT_1_layer_call_and_return_conditional_losses_21699) with ops with unsaved custom gradients. Will likely fail if a gradient is requested.

[kahst]

Let people use a separate test dataset instead of splitting from the training data when re-training

[joshctaylor]

Is it possible to get a description of how the species filter works, what are the datasources?

[Mattk70]

Is it possible to get a description of how the species filter works, what are the datasources?

Like this: https://github.com/kahst/BirdNET-Analyzer/discussions/234 ?

[conormcmahon]

This #234 is really interesting. One useful extension would be a simple tool to query that map and let us know whether a chosen point is represented with actual eBird data (and maybe how much by season?) or only with eBird filters. E.g. a script which takes a gps point and gives some feedback on how trustworthy the automated species might be.

[Mattk70]

Allow custom classifier export to Tensorflow Saved model format.

This would make it far easier to use the custom classifier in TFJS.