jealous1989/yolov5-tensorrt-1

A tensorrt implementation of yolov5: https://github.com/ultralytics/yolov5

yolov5-tensorrt

A tensorrt implementation of yolov5: https://github.com/ultralytics/yolov5

requirement

Please use torch==1.4.0 + onnx==1.6.0 + TRT 7.0+ to run the sample code
onnx-simplifier-0.2.9

The code

Add newly implemented upsample to get this working with current combination of onnx and tensorrt.

1. download weights file
2. modify yolov5*.yaml, replace nn.Upsample with Upsample
3. python main.py to run the benchmark
4. Generally, for image of size 640*640, using batchsize=1, the speedup is 4x on V100.

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
• Standalone infer script