12 May 2013 - Waterloo ON
I had the honor of presenting my Hyperopt and Skdata projects at Scipy2013 this year. The presentations were on filmed, and now they are up on youtube:
While I was at the Rowland Institute. David Cox encouraged me to pursue my goal of creating software to study computer vision algorithm spaces; Hyperopt and Skdata are two reusable Python libraries that came out of that. Hyperopt is a generic optimization framework for the sort of awkward search spaces that algorithm hyperparameters present. Skdata is a library of Python scripts for dealing with various benchmarking data sets, and for running correct, standard algorithm evaluation experiments on those data sets. More specialized software such as hyperopt-convnet uses hyperopt and skdata to run experiments on the generic image classification model that we presented at ICML 2013, and the workshop on representation learning
Citation:
J. Bergstra, D. Yamins, D. D. Cox (2013).
Hyperopt: A Python Library for Optimizing the Hyperparameters of Machine
Learning Algorithms.
Proc. SciPy 2013.
Citation:
J. Bergstra, N. Pinto, D. D. Cox (2013).
Skdata: Data Sets and Algorithm Evaluation Protocols in Python.
Proc. SciPy 2013.