Working on Deep Learning Training Performance

Contributed the NUMA Binding feature to open-source PyTorch, now available as an option in torchrun for improved training performance: https://docs.pytorch.org/docs/main/elastic/numa.html (in v2.10.0)

Example usage: torchrun --numa-binding=node --nproc_per_node=8 train.py

Worked with Prof. Kostas Daniilidis at the GRASP (General Robotics, Automation, Sensing & Perception) Laboratory on controlling an F-1/10th(http://f1tenth.org/) car using an Event Camera and RGB Camera with Deep Reinforcement Learning.

Worked on MLPerf with the Deep Learning Infrastructure Team (NSV)

Worked with Prof. Raj Rajkumar at Cylab, Carnegie Mellon University on Self-Driving Cars.

The work done is as follows:

1) Used the 3D graphics rendering engine called Ogre-3D (https://ogrecave.github.io/ogre/api/latest/index.html) to develop a simulation engine for incorporating the real-time environment to test the car.

2) The simulation engine which is called EMERALD-Emulator/Simulator for Embedded Real-time Autonomous Intelligent Driving, is used to allow the designer to test the car in various scenarios and environments to have all of it covered when an autonomous car moves in the real world.

3) Developed the road networks that can be produced on attaching the map of any city/town.

4) Added objects like trees, bushes to the engine and also came up with a Non-Dynamics version with which the car moves.

5) Developed some AI features like making the car respond to the traffic lights and stop signs under different scenarios.

Worked with Prof. Raj Rajkumar at CyLab,Carnegie Mellon University on Self Driving Cars

The work done is as follows:

1) VISION FRAMEWORK FOR NVIDIA DRIVE PX2: Worked extensively on the NVIDIA DRIVE PX2 to develop a Vision Framework that will be incorporated on the Cadillac SRX. I had used 4 Fish-eye cameras that was run on TEGRA-A and fed on to the Moving Object Detector and also to the OpenGL 3D View. My mentor and I did some test runs on a Zip car to test the framework. I had also devised a way to control 6 simultaneous camera capture of the two groups present on the DRIVE PX2,using socket programming. The method given by NVIDIA is only for a group that supports upto 4 cameras.

I had also contributed to the 3D view for the autonomous vehicles by helping the team setup the PX2 onto the car and also controlling the NVIDIA DRIVE PX2 to capture videos that was used by the team for calibration and annotation purposes.

I had also finished pipelining the live camera feed from NvMedia to the Moving Object Detector.

2) ENVIRONMENT MANAGER :Developed an algorithm to find out the position of the sun in order to aid the autonomous vehicles . This is to help the autonomous vehicle when the cameras are blinded by sunlight. I had also written a code snippet to get the real-time weather data from OpenWeatherMap in order to aid the navigation process in the autonomous vehicles.