Gain of EfficientNMS_TRT embed on TensorRT #6430
Description
Search before asking
- I have searched the YOLOv5 issues and found no similar feature requests.
Description
Hi YOLOv5 Community,
I wanna share with you our gains on TensorRT with a new pipeline to deal with YOLOv5, we embed the whole post-processing (namely the EfficientNMS_TRT plugin) into the Graph with onnx-graghsurgeon. The ablation experiment results are below. The first one is the result without running EfficientNMS_TRT, and the second one is the result with EfficientNMS_TRT embedded. As you can see, the inference time is even reduced, we guess it is because the data copied to the device will be much less after doing EfficientNMS_TRT. (The mean Latency of D2H is reduced from 0.868048 ms to 0.0102295 ms, running on Nivdia Geforce GTX 1080ti, using TensorRT 8.2 with yolov5n6 and scaling images to 512x640.) Check https://zhiqwang.com/yolov5-rt-stack/notebooks/onnx-graphsurgeon-inference-tensorrt.html for more details.
Use case
Deploy YOLOv5 quickly on TensorRT, and in this way, you do not need to write any C++ codes for post-processing.
Known cons
We have to update the TensorRT to 8.2 to call the EfficientNMS_TRT plugin. And seems that there is a bug about the float16 of this plugin: NVIDIA/TensorRT#1758 (comment).
Related Issue
Additional
And onnx-graphsurgeon is easy to install, you can just use their prebuilt wheels:
python3 -m pip install onnx_graphsurgeon --index-url https://pypi.ngc.nvidia.com
The detailed results:
[I] === Performance summary w/o EfficientNMS_TRT plugin ===
[I] Throughput: 383.298 qps
[I] Latency: min = 3.66479 ms, max = 5.41199 ms, mean = 4.00543 ms, median = 3.99316 ms, percentile(99%) = 4.23831 ms
[I] End-to-End Host Latency: min = 3.76599 ms, max = 6.45874 ms, mean = 5.08597 ms, median = 5.07544 ms, percentile(99%) = 5.50839 ms
[I] Enqueue Time: min = 0.743408 ms, max = 5.27966 ms, mean = 0.940805 ms, median = 0.924805 ms, percentile(99%) = 1.37329 ms
[I] H2D Latency: min = 0.502045 ms, max = 0.62674 ms, mean = 0.538255 ms, median = 0.537354 ms, percentile(99%) = 0.582153 ms
[I] GPU Compute Time: min = 2.23233 ms, max = 3.92395 ms, mean = 2.59913 ms, median = 2.58661 ms, percentile(99%) = 2.8201 ms
[I] D2H Latency: min = 0.851807 ms, max = 0.900421 ms, mean = 0.868048 ms, median = 0.867676 ms, percentile(99%) = 0.889191 ms
[I] Total Host Walltime: 3.0081 s
[I] Total GPU Compute Time: 2.99679 s
[I] Explanations of the performance metrics are printed in the verbose logs.
[I]
&&&& PASSED TensorRT.trtexec [TensorRT v8201] # trtexec --onnx=yolov5n6-no-nms.onnx --workspace=8096
[I] === Performance summary w/ EfficientNMS_TRT plugin ===
[I] Throughput: 389.234 qps
[I] Latency: min = 2.81482 ms, max = 9.77234 ms, mean = 3.1062 ms, median = 3.07642 ms, percentile(99%) = 3.33548 ms
[I] End-to-End Host Latency: min = 2.82202 ms, max = 11.6749 ms, mean = 4.939 ms, median = 4.95587 ms, percentile(99%) = 5.45207 ms
[I] Enqueue Time: min = 0.999878 ms, max = 11.3833 ms, mean = 1.28942 ms, median = 1.18579 ms, percentile(99%) = 4.53088 ms
[I] H2D Latency: min = 0.488159 ms, max = 0.633881 ms, mean = 0.546754 ms, median = 0.546631 ms, percentile(99%) = 0.570557 ms
[I] GPU Compute Time: min = 2.30298 ms, max = 9.21094 ms, mean = 2.54921 ms, median = 2.51904 ms, percentile(99%) = 2.78528 ms
[I] D2H Latency: min = 0.00610352 ms, max = 0.302734 ms, mean = 0.0102295 ms, median = 0.00976562 ms, percentile(99%) = 0.0151367 ms
[I] Total Host Walltime: 3.00591 s
[I] Total GPU Compute Time: 2.98258 s
[I] Explanations of the performance metrics are printed in the verbose logs.
[I]
&&&& PASSED TensorRT.trtexec [TensorRT v8201] # trtexec --onnx=yolov5n6-efficient-nms.onnx --workspace=8096
Are you willing to submit a PR?
- Yes I'd like to help by submitting a PR!