Developing agents and training models for CUDA

TensorRT

I researched deep-learning-based methods for interactive object understanding and object tracking in egocentric videos. I was advised by Assistant Professor Saurabh Gupta.

• Introduced a software-based kernel timing heuristic (DLSim) for neural network optimization

• Bridged C++ TensorRT and Python DLSim by implementing a server-client interface handling 10s of queries/s with convolutional layer parameter translation

• Compared 5 networks (ResNet-50, MobileNet, Inception-V4, etc), across 3 batch sizes and FP16, INT8 precision, meeting the <10% throughout reduction goal