nilomr/wytham-songtype-validation

Code to train a classifier to check the robustness of a manual classification of song types following McGregor & Krebs 1982

This repository contains the code to to train a classifier to check the robustness of a manual classification of Great Tit song types following McGregor & Krebs (1982). For more information on the preprocessing of the data, see this paper. Model training follows the steps described in this article.

A narrative code notebook including outputs can be found here.

Installation

1. Create a new environment, e.g. using miniconda:

conda create -n wytham-songtype-validation python=3.9

2. Clone this repository to your local machine, navigate to its root and install using pip:

git clone https://github.com/nilomr/wytham-songtype-validation.git
cd wytham-songtype-validation
pip install .

GPU installation

One of the steps to reproduce this example involves training a deep neural network, which requires compatible GPU resources.

If you want to reatrain the model, you will need a few more libraries that are not installed automatically with pykanto. The reason for this is that the are a bit finicky: which exact installation you need depends on which version of CUDA you have and the like.

I recommend that, if this is the case, you first create a fresh environment with conda:

conda create -n wytham-songtype-validation python=3.9

And then install torch, pykanto and this example including the extra libraries.

conda install -c pytorch pytorch torchvision   
pip install pykanto
# Navigate to the root of this repository, then:
pip install ."[torch]" # see the pyproject.toml file for other options

User guide

First, make sure that you have activated this project's environment (conda activate wytham-songtype-validation if you followed the instructions above). Then, navigate to /notebooks. This is where the scripts are located. They can all be run from the terminal, python <script-name>.

Expand user guide

Script	Description	Use
1_prepare-dataset.py	Ingests, creates spectrograms, and segments the dataset1	To run: python 1_prepare-dataset.py. Requires the output of this repository.
3_export-training-data.py	Exports the data required to train the deep learning model	python 3_export-training-data.py
4_train-model.ipynb	Model definition and training step	A separate, self-contained jupyter notebook. This is to make it easier to run interactively on a GPU-enabled HPC. If you don't want to retrain the model, you can skip this step.
5_save_labels.py	Saves the checked labels to a csv file	python 5_save_labels.py

_{© Nilo M. Recalde, 2021-present}

Footnotes

1. If you want to run this in a HPC you can use pykanto's tool for this, which makes it very easy (see Docs for more info). ↩