Object Detection with Transformers - From Training to Deployment

Explore object detection with transformers in this comprehensive 32-minute talk from Databricks. Learn the fundamentals of object detection, including key concepts and techniques, before delving into cutting-edge methods that utilize transformers to streamline the detection pipeline. Discover the main ideas behind DETR and Deformable DETR approaches, and gain an overview of Determined AI's deep learning platform capabilities, focusing on effortless distributed training. Master the process of training object detection models at scale and serving them using MLflow. Grasp essential topics such as mean average precision, transformer decoder architecture, positional encoding, and deformable attention. Follow along as the speaker demonstrates defining PyTorch trials, configuring experiments, and utilizing Determined's web UI for hyperparameter search, automatic fault tolerance, and model saving. By the end of this talk, acquire the knowledge to implement advanced object detection techniques and deploy them effectively in various applications, from medical image analysis to autonomous driving.

Introduction
Outline
Object Detection
Prediction Problem
Mean Average Precision
Why should we care
Transformer Decoder Architecture
Positional Encoding
Decoder
Prediction
Training
DebtR Performance
DebtR Drawbacks
Deformable Attention
Multiscale Features
Performance
State of the Art
Training Models
Defining PiTorch Trial
Defining Experiment Config
Determining Web UI
HP Search
Other Metrics
Tensorboard
Automatic Fault Tolerance
Save Model
Output
Results
Conclusion