aascode / pam_transferlearning

Passive Acoustic Monitoring and Transfer Learning

Transfer learning for passive acoustic monitoring

In this study we describe the implementation of transfer learning to create bioacoustic classification models as a means to reduce the complexity of implementing deep learning algorithms from scratch. We compared 12 modern CNN architectures across 4 datasets and focus on scarce data issues when attempting to create a passive acoustic binary classification model with as few as 25 verified examples. The datasets contained vocalisations of: the critically endangered Hainan gibbon (Nomascus hainanus), the critically endangered black-and-white ruffed lemur (Varecia variegata), the vulnerable Thyolo alethe (Chamaetylas choloensis), and the Pin-tailed whydah (Vidua macroura).

Our goal was to demonstrate that transfer learning is a suitable and accessible approach to implementing CNNs which require less human expert knowledge in machine learning.

This code accompanies the paper ``Passive Acoustic Monitoring and Transfer Learning''.

Each folder contains two main code (fine tuning and no fine tuning) examples for 25 input samples. Both will contain a notebook in the form of "Training...ipynb" and "Predicting_...ipynb". The former is the notebook to train the model and the latter to predict on test cases. The easiest approach is to try out the code on Google Colab (see below) as those notebooks automatically download a subset of data. Alternatively, to run the code locally, download the appropriate dataset (see below).

Authors

Emmanuel Dufourq, Carly Batist, Ruben Foquet and Ian Durbach

---

Requirements

Install all requirements using pip install -r requirements.txt

The following packages were used in development and have been tested on Ubuntu 20.04.3 LTS running Python 3.8 (NVIDIA-SMI 470.103.01, CUDA Version: 11.2, cudnn: 8.1.1)

• tensorflow==2.7.0
• SoundFile==0.10.3.post1
• scikit-learn==1.0.2
• scipy==1.8.0
• numpy==1.18.5
• jupyter==1.0.0
• pandas==1.4.1
• librosa==0.8.0
• yattag==1.14.0
• matplotlib==3.3.3
• Keras==2.7.0

---

Datasets

We provide a subset of data used in this study:

• critically endangered Hainan gibbon (Nomascus hainanus) - Link: https://zenodo.org/record/6328319#.YiSoBRuxVH4 DOI: 10.5281/zenodo.6328319
• critically endangered black-and-white ruffed lemur (Varecia variegata) - Link: https://zenodo.org/record/6331594#.YiTo4HUzZKs DOI: 10.5281/zenodo.6331594
• vulnerable Thyolo alethe (Chamaetylas choloensis) - Link: https://zenodo.org/record/6328244#.YiSo5BuxVH4 DOI: 10.5281/zenodo.6328244
• least concern pin-tailed whydah (Vidua macroura) - Link: https://zenodo.org/record/6330711#.YiSnuBuxVH4 DOI: 10.5281/zenodo.6330711

Executing code on Google Colab (easiest way to get started!)

• Thyolo alethe
  • Training (feature extractor frozen - 8 minutes): https://colab.research.google.com/drive/1KZDRrCEjzu3HzP3dLFmDbqXVO0eqEs7V?usp=sharing
  • Predicting (feature extractor frozen): https://colab.research.google.com/drive/1piBHAS5JX8bVdjBXnygejjF9fxmysla8?usp=sharing
  • Training (feature extractor fine-tuned - 8 minutes): https://colab.research.google.com/drive/1o1dmvaOWy6j3hdjAdtZrnV19jRto8Da6?usp=sharing
  • Predicting (feature extractor fine-tuned): https://colab.research.google.com/drive/1xKydHw2aoeVSP_JQtbVFCHni5i5c5pbc?usp=sharing
• Hainan gibbons
  • Training (feature extractor frozen): https://colab.research.google.com/drive/1oYO3UWTL1BfxZn0IA0BYcg8JMGnH980D?usp=sharing
  • Predicting (feature extractor frozen): https://colab.research.google.com/drive/1d-hGRPRIzHubwlTeedqoZCSX8RGp0r4q?usp=sharing
  • Training (feature extractor fine-tuned): https://colab.research.google.com/drive/1uj1OT6JZ7I9z0PRB3YgE11M3ikSaSufa?usp=sharing
  • Predicting (feature extractor fine-tuned): https://colab.research.google.com/drive/1HdnSbvBXyAkYZ4UoA4e-OUpLqeNWm-eV?usp=sharing
• Lemurs
  • Training (feature extractor frozen): https://colab.research.google.com/drive/1Nxa3hDdaR7nKpMdDGA0aBdalxMafRFvw?usp=sharing
  • Predicting (feature extractor frozen): https://colab.research.google.com/drive/1eknVT0M9CTHqLxX29gXi6eITodSJ7Edi?usp=sharing
  • Training (feature extractor fine-tuned): https://colab.research.google.com/drive/1yNv2GctC8B6Z0aIIdlglunctfeR6ZbW7?usp=sharing
  • Predicting (feature extractor fine-tuned): https://colab.research.google.com/drive/1YTn4o8klnFz_Kj9eHwyX4HoA_VVdrk0l?usp=sharing
• Pin-tailed whydah
  • Training (feature extractor frozen): https://colab.research.google.com/drive/1AIhBTQkaUkBrs9ADGU1r4fz1eRGKmyJ9?usp=sharing
  • Predicting (feature extractor frozen): https://colab.research.google.com/drive/1FeLYcKQlE29m64TTRWJ_suuHuI9sxREe?usp=sharing
  • Training (feature extractor fine-tuned): https://colab.research.google.com/drive/1KTOQlnUdY6Jv7wzYympA-PFOUgPqOOsX?usp=sharing
  • Predicting (feature extractor fine-tuned): https://colab.research.google.com/drive/14DkzYCnJXwdSuQno7ONrfX89G9UnvaV8?usp=sharing