Giters

dFuong / TLT-converter

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

This document captures simple instructions to run the TLT converter engine for the x86 platform

TensorRT version: 7.2.2-1

CUDA version: 11.1

CUDNN version: 8.0.5

Installation instructions:

You can follow the approprivate version TLT-converter or TAO-converter (TLT-upgraded) for Hardware: TLT-converter or TAO-converter

Check version TensorRT: dpkg -l | grep TensorRT

  1. Copy the executable to the target device

  2. Install openssl library using the following commam.

    sudo apt-get install libssl-dev

  3. Export the following environment variables

    export TRT_LIB_PATH=”/usr/lib/x86_64-linux-gnu"

    export TRT_INC_PATH=”/usr/include/x86_64-linux-gnu”

  4. Run the converter following the below guide: Sample usage

Note:

In Int8 mode, when using calibration table generated from the TLT v1.0 GA docker, if the version of the tensorrt on the deployment device is 5.x/6.x, please make sure to update the version number in the TensorRT calibration table file. Inorder to do so,

  1. Open the calibration.bin or calibration.txt file as generated by tlt-export using a simple text editor
  2. Update the first line in calibration file from TRT-5105-EntropyCalibration2 or TRT-6001-EntropyCalibration2 to TRT-7000-EntropyCalibration2.

You need change the first line TRT version siutable for enviroment TRT Hardware eg: TRT-7221-EntropyCalibration2 siutable for TensorRT-version

  1. Save the edited file with the new changes.
  2. Use this file with tlt-converter to generate the engine.
  3. Running TAO-converter as same as TLT-coverter.

Sample usage:

Inital running:
$chmod +x tlt-converter

Sample:
$./tlt-converter resnet34_peoplenet_pruned.etlt -k tlt_encode -c resnet34_peoplenet_int8.txt -o output_cov/Sigmoid,output_bbox/BiasAdd -d channel,height,width -i nchw -e peoplenet34_int8.engine -m $MAX_BATCH_SIZE -t $INFERENCE_PRECISION -b $BATCH_SIZE

Example:
$./tlt-converter peoplenet_v2/resnet34_peoplenet_pruned.etlt -k tlt_encode -c peoplenet_v2/resnet34_peoplenet_int8.txt -o output_cov/Sigmoid,output_bbox/BiasAdd -d 3,544,960 -i nchw -e peoplenet_resnet34_int8.engine -m 1 -t int8 

tlt_encode = KEY-TLT

After achived file engine, you can check inference:

Type: detectnet_v2,maskrcnn,unet Model-arch

Sample:
$tlt-infer type -e peoplenet_v2/infer_spec.txt -k tlt_encode -o infer_result -i file-image -v

About