Perch

A bioacoustics research project.

Directory of Things

We have published quite a few things which utilize this repository!

Perch (and SurfPerch!)

We produce a bird species classifier, trained on over 10k species.

• The current released Perch model is available from Kaggle Models.
• The current-best citation for the model is our paper: Global birdsong embeddings enable superior transfer learning for bioacoustic classification.
• The SurfPerch model, trained on a combination of birds, coral reef sounds, and general audio, is also available at Kaggle models. The associated paper is (as of this writing) available as a preprint.

The major parts of the Perch model training code is broken up across the following files:

• Model frontend - we use a PCEN melspectrogram.
• EfficientNet model
• Training loop
• Training launch script
• Export from JAX to Tensorflow and TFLite

Agile Modeling

Agile modeling combines search and active learning to produce classifiers for novel concepts quickly.

Here's Tutorial Colab Notebook we produced for Climate Change AI and presented at their workshop at NeurIPS 2023.

We maintain three 'working' notebooks for agile modeling in this repository:

• embed_audio.ipynb for performing mass-embedding of audio.
• agile_modeling.ipynb for search and active learning over embeddings.
• analysis.ipynb for running inference and performing call density estimation (see below).
• The code for agile modeling is largely contained in the inference directory, which contains its own extensive README.

The agile modeling work supports a number of different models, including our models (Perch and SurfPerch, and the multi-species whale classifier), BirdNet, and some general audio models like YamNet and VGGish. Adding support for additional models is fairly trivial.

Call Density

We provide some tooling for estimating the proportion of audio windows in a dataset containing a target call type or species - anything you have a classifier for.

Paper: All Thresholds Barred: Direct Estimation of Call Density in Bioacoustic Data

Code: See call_density.py and call_density_test.py. Note that the code contains some interactions with our broader 'agile modeling' work, though we have endeavoured to isolate the underlying mathematics in more modular functions.

BIRB Benchmark

We produced a benchmark paper for understanding model generalization when transferring from focal to passive acoustic datasets. The preprint is available here.

For details on setting up the benchmark and evaluation protocol, please refer to this brief readme with instructions. The evaluation codebase is in perch/chirp/eval.

To build the BIRB evaluation data, after installing the chirp package, run the following command from the repository's root directory:

poetry run tfds build -i chirp.data.bird_taxonomy,chirp.data.soundscapes \
    soundscapes/{ssw,hawaii,coffee_farms,sierras_kahl,high_sierras,peru}_full_length \
    bird_taxonomy/{downstream_full_length,class_representatives_slice_peaked}

The process should take 36 to 48 hours to complete and use around 256 GiB of disk space.

Source-Free Domain Adaptation and NOTELA

We have a paper on 'source-free domain generalization,' which involves automatic model adaptation to data from a shifted domain. We have a blog post where you can read more about it. The paper was published in ICML 2023. The code for this project has been archived. You can download a snapshot of the repository containing the code, which can be found in the chirp/projects/sfda directory.

Installation

We support installation on a generic Linux workstation. A GPU is recommended, especially when working with large datasets. The recipe below is the same used by our continuous integration testing.

Some users have successfully used our repository with the Windows Linux Subsystem, or with Docker in a cloud-based virtual machine. Anecdotally, installation on OS X is difficult.

You might need the following dependencies.

# Install Poetry for package management
curl -sSL https://install.python-poetry.org | python3 -

# Install dependencies for librosa
sudo apt-get install libsndfile1 ffmpeg

# Install all dependencies specified in the poetry configs
poetry install  --with jaxtrain

Running poetry install installs all Perch dependencies into a new virtual environment, in which you can run the Perch code base. To run the tests, use:

poetry run python -m unittest discover -s chirp/tests -p "*test.py"
poetry run python -m unittest discover -s chirp/inference/tests -p "*test.py"

Lightweight Inference

Note that if you only need the python notebooks for use with pre-trained models, you can install with lighter dependencies:

# Install inference-only dependencies specified in the poetry configs
poetry install

And check that the inference tests succeed:

poetry run python -m unittest discover -s chirp/inference/tests -p "*test.py"

Using a container

Alternatively, you can install and run this project using a container via Docker. To build a container using the tag perch, run:

git clone https://github.com/google-research/perch
cd perch
docker build . --tag perch

After building the container, to run the unit tests, use:

docker run --rm -t perch python -m unittest discover -s chirp/tests -p "*test.py"

This is not an officially supported Google product.