Multimodal Biosensing

Requirements

To install requirements:

pip install -r requirements.txt

Training

You can train each stage separately (the encoders, the single-variable prototypes, and multivariable prototypes).

1. python3 -m src.train.encoding.train --dataset <dataset> to train the univariate encoders.
2. python3 -m src.train.single_variable_prototypes.kmeans_silhouette --dataset <dataset> to compute the optimal number of single-variable prototypes for each variable in your data.
3. python3 -m src.train.single_variable_prototypes.train --dataset <dataset> --initialize to initialize the single-variable prototypes with k-means++ and then train them.
4. python3 -m src.train.multivariable_prototypes.train --dataset <dataset> to train the multivariable prototypes.

Once you've determined your hyperparameters, you can run the pipeline end-to-end: python3 -m src.eval.end_to_end --dataset <dataset> --resamples <resamples>

Our learned models are located in src/models and are organized by dataset. src/models/comparisons contains saved models for our ablation studies on our simulated dataset.

Evaluation

To evaluate the fully trained model on the test set:

python -m src.eval.eval --dataset <dataset>