Query on DAT_PATH Configuration for TensorRT Engine Build in uniad-trt Inference_app Step

[Josh-Stick]

Thank you very much for your excellent work. I encountered some problems while building uniad-trt. I would like to know what the DAT_PATH=<path_to_dumped_inputs> in the command for building the TensorRT engine in the inference_app step refers to. How should I get this?

[Jian-danai]

See https://github.com/NVIDIA/DL4AGX/blob/master/AV-Solutions/uniad-trt/tools/export_onnx.py#L250
DAT_PATH=<path_to_dumped_inputs> is <path_to_uniad-trt>/UniAD/dumped_inputs

BTW, for the engine build command, if trtexec inference failed, please try OPT=905 in

MIN=901
OPT=901
MAX=1150
DAT_PATH=<path_to_dumped_inputs>
SHAPES="prev_track_intances0:${MIN}x512,prev_track_intances1:${MIN}x3,prev_track_intances3:${MIN},prev_track_intances4:${MIN},prev_track_intances5:${MIN},prev_track_intances6:${MIN},prev_track_intances8:${MIN},prev_track_intances9:${MIN}x10,prev_track_intances11:${MIN}x4x256,prev_track_intances12:${MIN}x4,prev_track_intances13:${MIN}"
INPUTS="max_obj_id:${DAT_PATH}/max_obj_id.dat,img_metas_can_bus:${DAT_PATH}/img_metas_can_bus.dat,img_metas_lidar2img:${DAT_PATH}/img_metas_lidar2img.dat,img:${DAT_PATH}/img.dat,use_prev_bev:${DAT_PATH}/use_prev_bev.dat,prev_bev:${DAT_PATH}/prev_bev.dat,command:${DAT_PATH}/command.dat,timestamp:${DAT_PATH}/timestamp.dat,l2g_r_mat:${DAT_PATH}/l2g_r_mat.dat,l2g_t:${DAT_PATH}/l2g_t.dat,prev_track_intances0:${DAT_PATH}/prev_track_intances0.dat,prev_track_intances1:${DAT_PATH}/prev_track_intances1.dat,prev_track_intances3:${DAT_PATH}/prev_track_intances3.dat,prev_track_intances4:${DAT_PATH}/prev_track_intances4.dat,prev_track_intances5:${DAT_PATH}/prev_track_intances5.dat,prev_track_intances6:${DAT_PATH}/prev_track_intances6.dat,prev_track_intances8:${DAT_PATH}/prev_track_intances8.dat,prev_track_intances9:${DAT_PATH}/prev_track_intances9.dat,prev_track_intances11:${DAT_PATH}/prev_track_intances11.dat,prev_track_intances12:${DAT_PATH}/prev_track_intances12.dat,prev_track_intances13:${DAT_PATH}/prev_track_intances13.dat,prev_timestamp:${DAT_PATH}/prev_timestamp.dat,prev_l2g_r_mat:${DAT_PATH}/prev_l2g_r_mat.dat,prev_l2g_t:${DAT_PATH}/prev_l2g_t.dat"
LD_LIBRARY_PATH=<path_to_TensorRT>/lib:$LD_LIBRARY_PATH \
<path_to_TensorRT>/bin/trtexec \
  --onnx=<path_to_ONNX> \
  --saveEngine=<path_to_engine> \
  --plugins=<path_to_libuniad_plugin.so> \
  --verbose \
  --dumpLayerInfo \
  --dumpProfile \
  --separateProfileRun \
  --profilingVerbosity=detailed \
  --useCudaGraph \
  --minShapes=${SHAPES//${MIN}/${MIN}} \
  --optShapes=${SHAPES//${MIN}/${OPT}} \
  --maxShapes=${SHAPES//${MIN}/${MAX}} \
  --loadInputs=${INPUTS}

BTW, please wait for our TensorRT-10 support for this project if you do not have access to TensorRT-8.6.13