Ordinal Triplet Loss

PyTorch implementation of the ordinal triplet loss function detailed in the Interspeech paper.

Background

Ordinal triplet loss (OTL) is a classification loss function that accounts for the relative distance between ordinal classes. The two key techniques used to realize this are soft labels and a modified triplet loss function.

Quick Start

• To use soft labels only, see the mk_y_slabs and slab_predict functions in slab/utils.py. For usage examples, see slab/main.py.

• To use the ordinal triple loss function, see the OrdinalTripletLoss class in otl/utils.py. For usage examples, see otl/main.py.

Reproducing Results

• Download the challenge data, create a directory called data at the same level as baseline, and move labels.csv and the features directory into data
• The baseline directory contains instructions on how to run the baseline SVR and MLP models
• The slab directory contains instructions on how to run an MLP trained with soft labels
• The otl directory contains instructions on how to run an MLP trained using ordinal triplet loss