Skip to content

rentzi/sparseRegularizers

Repository files navigation

sparseRegularizers

Generative model with different thresholding functions based on an upcoming paper

Getting started

install jupyter, numpy, and matplotlib

follow the instructions and run 0 download preproces and whiten images to download the natural images from the Hateren Schaaf database, to preprocess them, and to create and store input image patches for the generative algorithm

change the variables accordingly and run 1 Generative model with different regularizers to process the input image patches with the generative model (create sparse codes and learn the overcomplete basis) with one of the four thresholding functions (ISTA, Hard thresholding, Half thresholding, CEL0)

This implementation of the generative models uses a learning model where the learning rates decay with iterations based on a time-based decay schedule, with decay rates for both learning rates being their initial values divided by 50. We used this implementation for CEL0 for dictionary sizes greater than 500 units. For the other methods the learning rates were the same across iterations. If you want to run the later implementation set the vectors lr_rV and lrPhiV in generativeModels.py to fixed values

About

Generative model with different thresholding functions

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published