2022年GTC大会演讲资料：NVIDIA 开发用于视觉任务的传感器.pdfVIP

DEVELOPING INFERENCE-EFFICIENT TRANSFORMERS FOR VISION TASKS JOHN YANG, LE AN (JOINT WORK WITH ANURAG DIXIT, JINKYU KOO, AND JOSH PARK) DEVELOPING INFERENCE-EFFICIENT TRANSFORMERS FOR VISION TASKS AGENDA ▪ Preliminary: transformer and transformer in vision ▪ Bottlenecks in current transformers ▪ Design of efficient transformer for depth estimation ▪ Benchmark and deployment with TensorRT ▪Generalization to semantic segmentation ▪Summary (VISION) TRANSFORMERS TRANSFORMER Attention is All You Need1 ▪ A deep learning mo