mortiniera / hogwild-multicore

Multicore implementation of the sparse SVM problem

Authors

• Timoté Vaucher
• Patrik wagner
• Thevie Mortiniera

Hogwild-multicore

Welcome to this implementation of the sparse SVM problem described in the original Hogwild ! paper using the numerous core of a computer. This was implemented during the spring of 2019 in the CS-449 Systems for Data Sciences at EPFL. This is the second milestone of a project including a synchronous and asynchronous distributed version of Hogwild implement in Spring 2018 and a Spark version implemented in the first milestone.

Prerequisites and Setup

We suppose that you have Anaconda installed on your computer in order to create a conda environment and that you downloaded the datasets on your machine (if you want to use the scipy version of hogwild-multicore, otherwise python 3.7 is enough)

conda create -n hogwild-multicore python=3.7 numpy scipy
source activate hogwild-multicore

Run

You may need to adapt the paths in settings.py to suit your architecture then from the src folder :

python hogwild.py

You can look at the help using python hogwild.py -h

Parameters

• -l or --lock Use a lock version of Lock-free Hogwild to avoid concurrent update
• -n or --niter To select the number of iterations of SGD default : 400
• -p or --process To select the number of process to run onto default : os.cpu_count
• -vor --verbose To show the train and accuracy loss at train time
• --notest To run only the training step and not load the test set / compute test accuracy