Introducing Mouffet, a unified framework to make model creation easier and more reproducible

Abstract

1. Biological and ecological models are being increasingly used to explain the natural world. Model creation is an iterative process requiring two steps: training and evaluating the models. However, this process can become complex when multiple models are trained and evaluated at the same time. Besides, development steps can be lost, reducing the reproducibility of model creation. 2. We introduce Mouffet, an open-source Python framework that aims to make model creation easier, more robust, and more reproducible. It provides a set of configuration files and high-level Python interfaces that help managing data, training, and evaluating models. To improve reproducibility, every step of the model creation process, including the options used, are saved. 3. Mouffet introduces the notion of scenarios that allow users to define multiple training or evaluation tasks in a single configuration file. This not only facilitates model creation but enables users to define experimental plans to study the effect of selected parameters on training or evaluation. 4. While initially developed for deep learning models, Mouffet is independent of the implementation of the models. Therefore, it could be successfully used to compare different modelling approaches. Besides, its ease of use makes it a choice tool for ecologists, even when not familiar with complex model creation.

Deep learning for passive acoustic monitoring: how to study changing phenology in remote areas

“…Finally, while BioSoundNet is currently targeted to bird detection, its core inputs are audio files, allowing great versatility across many vocal species, as long as the relevant training data is provided. The adaptable architecture of BioSoundNet facilitates straightforward retraining (Christin, Hervet, and Lecomte 2019;Christin and Lecomte 2022), rendering it a valuable resource for ecologists engaged in acoustic recording studies involving various taxa. And while we used a community approach here, it could also be used to quickly detect targeted species, depending on the examples provided in the training dataset.…”