In this page, PeopleNet’s fps on Xavier is claimed as 314FPS for Resnet-34.
But I tried to test using DeepStream.
Tested with 8 video stream from mp4 video files, FPS is less than 5FPS for batch size=8.
May I know why so much difference?
Please note that the figure is showing the result of int8 model @bs16. Also it is a pruned model. And the numbers in the table are inference performance measured using the trtexec tool in TensorRT samples.
So, please download the int8 pruned model and generate trt engine, then run trtexec. I can get similar fps result.
$ sudo nvpmodel -m 0
$ jetson_clocks
$ wget https://api.ngc.nvidia.com/v2/models/nvidia/tlt_peoplenet/versions/pruned_quantized_v2.1/files/resnet34_peoplenet_pruned_int8.etlt
$ wget https://api.ngc.nvidia.com/v2/models/nvidia/tlt_peoplenet/versions/pruned_quantized_v2.1/files/resnet34_peoplenet_pruned_int8_gpu.txt
$ tlt-converter resnet34_peoplenet_pruned_int8.etlt -k tlt_encode -o output_bbox/BiasAdd,output_cov/Sigmoid -d 3,544,960 -m 16 -t int8 -c resnet34_peoplenet_pruned_int8_gpu.txt -e resnet34_peoplenet.engine
$ /usr/src/tensorrt/bin/trtexec --loadEngine=resnet34_peoplenet.engine --int8 --batch=16 --useSpinWait
This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.