This README provides code for training, testing, and deploying, different convolutional neural network model architectures for automated detection and classification of acoustic data.
Users can train both binary and multi-class classification models on spectrogram images, and evaluate their performance on test datasets. The package includes tools for performance evaluation, allowing easy identification of the best-performing model. Once the best-performing model is identified, it can be deployed for large-scale inference on multiple sound files. In addition to classification, trained CNNs can be used as feature extractors. Combined with unsupervised clustering, this enables visualization of differences in acoustic signals.
A detailed usage guide can be found at: https://denajgibbon.github.io/gibbonNetR/
This package assumes a basic understanding of machine learning, deep learning, and convolutional neural networks (CNNs). Users should be familiar with training and evaluating models, as well as handling spectrogram image data. For those new to these concepts, we recommend reviewing foundational machine learning and deep learning resources before using this package. a good starting point would be:
Stowell, Dan. “Computational bioacoustics with deep learning: a review and roadmap.” PeerJ 10 (2022): e13152. https://peerj.com/articles/13152/
Some practical information on improving model performance can be found here: https://kahst.github.io/BirdNET-Analyzer/best-practices/training.html#
You can install the gibbonNetR package from its repository using devtools:
# If you don't have devtools installed
install.packages("devtools")
# Install gibbonNetR
devtools::install_github("https://github.com/DenaJGibbon/gibbonNetR")
# The first time you use the package 'torch' will need to install additional packages. You can start the process using the following:
library(torch)
library(gibbonNetR)
# Set file path to spectrogram images
filepath <- system.file("extdata", "multiclass/", package = "gibbonNetR")
# Train simple CNN model
train_CNN_multi(
input.data.path = filepath,
test.data = paste(filepath, "/test/", sep = ""),
architecture = "alexnet", # Choose 'alexnet', 'vgg16', 'vgg19', 'resnet18', 'resnet50', or 'resnet152'
unfreeze.param = TRUE,
batch_size = 6,
class_weights = rep((1 / 5), 5),
learning_rate = 0.001,
epoch.iterations = 1, # Or any other list of integer epochs
early.stop = "yes",
save.model = FALSE,
output.base.path = file.path(tempdir(), "/MultiDir/", sep = ""),
trainingfolder = "test_multi",
noise.category = "noise"
)