Getting Started

In this section, the example cnn_cost_example from Examples is explained to see how SpArSeMod works.

This is just a tutorial showing how to use SpArSeMod, for the specifics, background theory and related, reading the paper SpArSe: Sparse Architecture Search for CNNs on Resource-Constrained Microcontrollers is highly recommended.

If you want to skip the paper and read a summary of most important facts, you can go to [./]

Following cnn_cost_example

In every run we are going to need 4 new components written: search space, network builder, main call and configuration. The use of morphisms is optional but recommended.

Clone the repository and follow installation instructions. You will need anaconda or install yourself the dependencies.

Getting the data

The tutorial is based on a modified version of CIFAR10, the CIFAR10 Binary, where instead of 2 classes we have 2. First run the file examples/download_data.py it will download CoST (Corpus of Social Touch) and create the necessary directories. Then, download the CIFAR Binary from this link and extract the files cifar10binary.testand cifar10binary.train into the examples/data/data_cifar2 folder.