KTH trains Deep Learning models on the Qarnot platform

As a member of the KTH Formula Student team, we are building a self-driving open-wheeled car to participate in the FSG 2020. One of the important aspects of building such a car is the perception of the environment, e.g. detecting obstacles, people, roads, etc.

For the competition, we needed to detect traffic cones of different colors and sizes. Deep learning (DL) solutions in perception have been extremely successful in the recent past. Therefore, we decided to create our own deep learning solution to detect cones. Our solution is based on the YOLOv3 neural network architecture. To improve performance, this network must be trained to reliably detect cones signifying position and color.

The biggest challenge in developing such a solution is having a large-scale computing resource for training, and our general purpose laptops are not good enough. This is when we use Qarnot services.

Qarnot provided us with the most advanced hardware and a good size of RAM. Qarnot allowed us to run our solution in the cloud in a Docker container. One of the benefits is that once the system is configured, we don't have to worry about new hardware and drivers. Downloading data was really smooth with their web interface.

A deep learning model requires many iterations to train the model with different configurations to get the best performance. With Qarnot services, these iterations were fast for us because we could train our model faster and it saved us a lot of time. Also, we can access intermediate results and training progress from anywhere and we don't have to be in a physical location. This makes training more convenient.