Distributed Support Vector Machine trained with Stochastic Gradient Descent using Kubernetes (HOGWILD!)
Brune Bastide, Augustin Prado, Romain Choukroun
This is an example of how to run a Support Vector Machine trained using Stochastic Gradient Descent distributed using Kubernetes/Docker. The current version of the project can be run in synchronous or asynchronous modes, which are explained in the report, see the report folder to learn more !
The project exclusively runs in Kubernetes. You will need to create a persistent volume and change the claim name in the Kubernetes config file to fit it.
To have a full view of what happens in Kubernetes, on the server and two of the clients clients, change in the config file the TERMINAL variable by the name of your terminal, and run logging interface: ./logging_interface.sh. You can setup your experiment in the config file, and when you're done, simply run: ./run_kub.sh which will run the project in Kubernetes. You can follow the progress in the logging interface. If you have not uncommented the automatic termination in the run_kub.sh file, you can simply run: ./del_kub.sh to terminate the experiment gracefully.
Running the code on 20 000 samples in the training set with 5 clients and a mini batch of size 16 takes less than a minute and returns an accuracy of 95%.
To learn more about the project, do not hesitate to checkout our report in the report folder !