yolov5-tensorrt
A tensorrt implementation of yolov5: https://github.com/ultralytics/yolov5
requirement
Please use torch>=1.6.0 + onnx>=1.6.0 + TRT 7.1+ (fix upsample issue) to run the sample code
onnx-simplifier-0.2.16
The code
Add newly implemented upsample to get this working with current combination of onnx and tensorrt.
0. prepare above mentioned environment.
- git clone && git submodule update --init
- download weights file (use yolov5/models/export.py)
- python main.py to run the benchmark
- Generally, for image of size 640*640, using batchsize=1, the speedup is 4x on V100.
Updates
- 20201004 update to track yolov5 - v3.0 release. download model file from official websites please.
TODO
- NMS support
- dynamic shape or dynamic batchsize support (won't implement soon because onnx-simplifier only supports fixed shape)
- FP16 numerical issue and performance investigation
- Benchmark