Home

Welcome to the wiki of Calvin's Data Science Toolbox (CDST)!

Overview

The novelty of the CDST package comes in two-folds:

• Architectural Hyperparameter (AH) for Neural Network
• Reduced/Aggregated Output Resolution Regression

CDST contains a set of pre-designed deep learning algorithms based on Pytorch, these include:

• General Scalable Deep Learning Fully Connected Network (DNN)
• Calvin's Scalable Parallel Downsampler (CSPD)
• Ordinal Hyperplane Loss Classifier (OHPL) (Reference publication)

The Architectural Hyperparameter (AH) sampling can be used for all of the above algorithms and the Reduced/Aggregated Output Resolution Regression is designed specifically for CSPD. Demo Jupyter notebook are provided for applying these pre-designed deep learning algorithms with ready to use data object and ready to use data partitioning methods for performing K-fold cross-validation over hyperparameter sets. Using these pre-designed object and methods, the hyperparameter tuning can be done using the hyperparameter tuning library Ray Tune and evaluated statistically.