faunanet

faunanet - A bioacoustics platform based on birdnetlib using neural networks

What is faunanet?

faunanet is an extension of Birdnet-Analyzer, and uses birdnetlib as its basis. faunanet was developed with the goal to provide a platform for bioacoustics research projects, started at the Interdisciplinary center for scientific computing at the University of Heidelberg.

Features

Using the birdnetv2.4 model by default, faunanet provides three core features:

• Easily and arbitrarily exchange the underlying maching learning model for bioacoustics.
• Easy configuration using YAML files which are stored alongside the analysis results.
• Integrated, extendible REPL with which to interact with which to interact with a running instance.

The main element of faunanet is a 'watcher' class that continuously monitors a folder for incoming data files and allows for on-the-fly model- or parameter-change via the REPL. It can also be used as a library in your python project. By extending elements of bridnetlib rather, faunanet conserves the latter's capabilities while still adding its own functionality on top of it.

Getting started

Please refer to the 'Getting faunanet up and running' section in the documentation for an introduction.

Bugs, issues and feature requests

Please use the issue tab of the github page to report any bugs or feature requests.

Contributions

Feel free to contribute to this project by opening a pull request here.

Usage

Installation

Make sure that ffmpeg is installed on your system:

# for debian based linux
sudo apt install ffmpeg

or for macOS using homebrew:

brew install ffmpeg

On windows, you can use chocolatey or other package managers to the same effect:

choco install ffmpeg

ffmpeg is needed for the audio preprocessing that is done by faunanet when analyzing your data.

You can install faunanet from pip with

pip install faunanet[option]

with option being "tensorflow" for the full tensorflow suite, or "tensorflow-lite" to install only the tflite_runtime package which only provides restricted support for model operations, while full tensorflow supports all models built with it. Installing it from pip will install the newest release version of faunanet. One of the two is needed to run faunanet, with tensorflow being the appropriate one for most usecases.

You can also install it from source to get the latest development chagnes. To do this, clone the repository, create a python virtual environment, and install faunanetinto it:

    cd path/to/dir/where/faunanet/should/live 

    git clone https://github.com/ssciwr/iSparrow.git

    python3 -m venv ./path/to/faunanet/venv

    source ./path/to/where/venv/bin/activate 

    cd ./faunanet 

    python3 -m pip install .[tensorflow]  # or .[tensorflow-lite] for tflite models

On windows, the line source ./path/to/faunanet/venv/bin/activate must be replaced with python3 ./path/to/faunanet/venv/scripts/activate. You can also select which installation options you want in the last line, either tensorflow or tensorflow-lite. Additionally, if you want to do further development, the installation must be modified by replacing the install command, adding 'editable' mode and development dependencies.

    python3 -m pip install  -e .[tensorflow, dev]

For building the documentation locally, you have to install with yet another set of dependencies:

Setup

faunanet needs three destinations on your filesystem in order to run:

• a directory where the machine learning models for sound classification should be stored
• a directory where incoming data files are stored.
• a directory where the analysis results should be stored

faunanet has a method set_up for this purpose, that is called with the path to a .yml configuration file. This file contains all the necessary information to set up a working faunanet installation. This method must be called upon first usage in order to have a fully functioning faunanet installation that documents its state correctly upon usage. Below is an example of the structure faunanet demands:

Directories: 
  home: ~/faunanet
  data: ~/faunanet/data
  models: ~/faunanet/models 
  output: ~/faunanet/output

You can copy the above into a .yml file and customize the paths to whatever you want them to be. The ~ will be automatically expanded to the path to your home directory. The method can be used as follows in python code

    from faunanet import faunanet_setup as sps 
    filepath = Path("path", "to", "the", "install_config.yml")

    sps.set_up(filepath)

It is important to have an outermost node called Directories. Aside from creating the directories named in the installation config file, the installation method will download the default tensorflow models from hugginface and will create faunanet directories in you OS default config and cache folders. On Linux, these would be ~/.config and ~/.cache, respectively.

In the repl, this would work like this:

   faunanet # enter repl 

   set_up --cfg=./path/to/custom/install_config.yml 

This does the same as the call to sps.set_up call above, so you can refer to the documentation of that function for further information, too.

Using faunanet as a library

faunanet can be used as a library in your own application. Just add

import faunanet 

to have it available in your module, and make sure you use the correct virtual environment in which faunanet is installed. Keep in mind that faunanet in itself extends birdnetlib, so make sure to check the latter's documentation, too.

Running faunanet

faunanet provides its own, small, REPL for interacting with a running instance. This can be used to start, stop, pause or continue it, to change classifier models or to query it's current state, input and output folders and so on. To get an overview over the available commands, you can just type faunanet in a terminal with the virtual environment being activated that faunanet has been installed into. Alternatively, refer to the documentation.

Using the docker image

Run the docker-hub image

You can also run faunanet in docker by pulling the latest faunanet image from docker-hub and running it via terminal command:

docker run -ti --rm \ 
     -v /path/on/host/for/faunanet/configs:/root/faunanet/config \
     -v /path/on/host/for/faunanet/output:/root/faunanet/output \ 
     -v /path/on/host/for/faunanet/models:/root/faunanet/models \ 
     -v /path/on/host/for/faunanet/data:/root/faunanet/data \ 
     mahawo/faunanet_{OPTION}:latest

Of special interest are the mounted volumes, i.e., the paths behind the -v arguments:

• first: for config files
• second: for analysis output
• third: for models
• forth: for incoming data. If you run the system via docker compose (see below) in conjunction with faunanet-record you do not need this, because faunanet-record will take care of this folder.

{OPTION} corresponds ot tf for tensorflow or tflite for tensorflow-lite.

Built the image yourself

To built the dockerfile that comes with the package yourself you can use the following docker command:

docker build --build-arg INSTALL_OPTION=TENSORFLOW_OPTION \
    -t your-dockerhub-username/your-image-name:tag \
    -f path/to/Dockerfile .

Where the TENSORFLOW_OPTION has to be replaced with either tensorflow or tensorflow-lite. The dockerfile itself is very simple and can be modified to your liking.

FROM python:3.11-slim

WORKDIR /root

RUN apt-get update && apt-get install --no-install-recommends -y ffmpeg -y --no-install-recommends && apt-get clean && rm -rf /var/lib/apt/lists/*

# add install option 
ARG INSTALL_OPTION

# install with the necessary option
RUN pip install faunanet[${INSTALL_OPTION}]
WORKDIR /root

RUN mkdir /root/faunanet/config 

# add entrypoint
CMD ["faunanet"]

In order to build an image for the ARM64 architecture often used by raspberry PI or other edge devices, you can use docker buildx in conjunction with qemu (tested on linux machine):

 docker buildx build --platform=linux/arm64 -t containername:tag -f ~/path/to/docker/file/dockerfile.dockerfile . --push

Have a look here for more info.

Using faunanet with other services via docker compose

faunanet can be run together with faunanet-record using docker compose or together with other containers of your choice. You can use the following docker-compose file as a starting point, which also comes with the installation:

services:
  faunanet:
    image: mahawo/faunanet:latest
    tty: true 
    stdin_open: true
    volumes:
      - ~/faunanet/config:/root/faunanet/config
      - ~/faunanet/output:/root/faunanet/output
      - ~/faunanet/models:/root/faunanet/models
      - ~/faunanet/data:/root/faunanet/data
    environment:
      - RUN_CONFIG=analysis_config.yml
  faunanet_record:
    image: mahawo/faunanet_record:latest
    tty: true 
    stdin_open: true
    volumes:
      - ~/faunanet/config:/root/faunanet/config
      - ~/faunanet/data:/root/faunanet/data
    devices:
      - /dev/snd:/dev/snd # this needs to be the microphone device
    environment:
      - RUN_CONFIG=record_config.yml

To locate the files from an existing pip installation, use the following python script, or pull them from the docker directory in faunanet home directory after it has been set up (see Setup). The environment variables RUN_CONFIG for each service here can hold the name of config files that are stored in the directory mounted into /root/faunanet/config.

When working with config files for the docker version, make sure to refer to the folders within the docker container and not the local ones. Otherwise, the system will fail with a