Torched and Fried Documentation

Torched and Fried (torchfry) is a software package that implements kernel approximation algorithms Fastfood (Le, Sarlos, Smola, 2013) and Random Kitchen Sink (Rahimi, Recht, 2007) as custom PyTorch layers for use in different networks.

Our motivation for implementing the Fastfood algorithm with PyTorch is to create an easy to use package that makes kernel methods scalable to high-dimensional datasets through efficient random feature layers like that in Deep Fried Convnets (Yang, Moczulski, Demil, et al., 2014). Through testing, we have found training times of Fastfood networks to be faster than Random Kitchen Sink networks. The improvement is about 20% decrease in time. The parameter count of Fastfood networks is much lower than those for Random Kitchen Sink networks, all while retaining the same image classification performance.

Table of Contents

• torchfry.networks
• torchfry.transforms