BNNBaseline

ResNet-50 BNN class

Contains:

• Dataset partitioning as defined in paper [to be updated to URL]
• Feature extraction to extract log-mel filterbank features
• Model training and evaluation
• Uncertainty quantification and meta-analysis

Code structure:

• Code configuration in lib/config.py
• Feature extraction in lib/extract_feat.py
• Model training loop in lib/train_model.py
• Model evaluation and meta-analysis in lib/evaluate_model.py

Requirements:

This module is designed to be run with the dockerfile supplied in the main README. If you wish to skip the docker file installation and only use this package, the code was developed with the conda environment on AWS of pytorch_latest_p36 with an additional command of:

conda install -c conda-forge -y librosa

How to configure code:

• Select the parameters of the feature transform you wish to use in lib/vggish/vggish_params.py.
• You may also bypass this and implement your own features directly in lib/extract_feat.py.
• Select the output directories, and model hyperparameter options in lib/config.py.

About the model

Introduction

ResNet-50 has seen success in many computer vision tasks, and has shown state-of-the-art performance in related acoustic methods acoustic mosquito detection, heart murmur detection.

Feature extraction

The feature extraction pipeline is imported in lib/vggish from torch vggish, which is a port from Audioset.

Training loop

The model is trained and early stopped according to the pre-defined splits of train, val and test for each carefully selected data partition. These are created upon feature extraction in extract_feat.py. Please refer to the papers [URL here] and [here] for a thorough description of the reasoning to create these data partitions. By default, we use the BCELoss() criterion with the Adam optimiser with the learning rate set in config.py.

Bayesian Neural Network modification

We modify the final layers for compatibility with our data in lib/ResNetSource.py Furthermore, we have augmented the construction blocks BasicBlock() and Bottleneck(), as well as the overall model construction, to feature dropout layers to act as an approximation for the model posterior at test-time. Dropout is implemented implicitly in ResNetSource.py, to not interfere with the behaviour of model.eval(), which by default disables dropout layers at test time, removing the necessary stochastic component. The final self.fc1 layer is of size [2048,N], where N is the number of classes for the cross-entropy loss function, or 1 if used with the binary cross-entropy loss (default). A quick way to check the requirement is to print x.shape() before the creation of the fc1 layer.

Uncertainty quantification and evaluation

To help understand the role of predictive entropy and mutual information in the outputs of a Bayesian neural network, I strongly recommend reading How do we go about analysing uncertainty?