search

HaohaoNJU / CenterPoint

TensorRT deployment for CenterPoint Lidar Detection Model.
MIT License
270 stars 56 forks source link

The engine plan file is generated on an incompatible device, expecting compute 7.0 got compute 8.6, please rebuild. #3

Closed hufangjian closed 2 years ago

hufangjian commented 2 years ago

I use V100，please provide the origin model ，thank you

HaohaoNJU commented 2 years ago

Hi, the baseline model is pretrained by Tianwei, you may contact him to get access to the ckpt, yintianwei@utexas.edu .

Natsu-Akatsuki commented 2 years ago

@hufangjian Hi, I work around this problem to use the trtexec to rebuild the engine from your nvidia environment. e.g.: $ trtexec --onnx=pfe_baseline32000.onnx --workspace=16384 --saveEngine=pfe_baseline_fp16.engine $ trtexec --onnx=rpn_baseline.onnx --workspace=16384 --saveEngine=rpn_baseline_fp16.engine. b.t.w. the path(like model path) in the source code is hard coded, you need to make some modification.

Natsu-Akatsuki commented 2 years ago

On the other hand, you could set the bool value params.load_engine = false in file samplecenterpoint.cpp