Train and serve object detectors for autonomous driving
Who is this presentation for?
- ML practitioners and researchers in the autonomous driving industry
The object detector is a key component in the perception system in autonomous driving vehicles. With more and more data collected continuously, there have been growing demands on fast and efficient training to speed up the development cycle. On the other side, inference on car is latency sensitive task, requiring detector to be high-optimized under limited resources.
Pengchong Jin walks you through a typical development workflow on GCP for training and deploying an object detector to a self-driving car. He demonstrates how to train the state-of-the-art RetinaNet model fast on the COCO (or Waymo) dataset using Cloud TPUs, and scale up the model effectively by leveraging large-scale Cloud TPU Pods. And he explains how to use the TensorFlow GPU API to export a Tensor-RT optimized model for GPU inference that can run in the car in real time.
You’ll discover how to set up a GCE environment, walk through RetinaNet code, launch training on Cloud TPU Donut, visualize model results using Colab, scale up training using Cloud TPU Pod, export TensorRT optimized saved model using TensorFlow GPU API, and learn further performance tuning.
- A basic understanding of TensorFlow and object detection
What you'll learn
- Learn why GCP and Cloud TPU and GPU is the one-stop destination for what you need for model training and serving
Pengchong Jin is a senior software engineer on the TensorFlow-E2E team at Google Brain, focusing on computer vision model development. He works closely with various autonomous driving companies on delivering object detection E2E solution on TPU and TensorRT inference. Previously, he worked on developing the internal object detector to serve various Google products, including photos, lens, image search.
For conference registration information and customer service
For more information on community discounts and trade opportunities with O’Reilly conferences
For information on exhibiting or sponsoring a conference
For media/analyst press inquires