Request: Pre-built TensorRT Engine for PeopleNet ResNet34 INT8 (A100 + TensorRT 10.13)

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Issue Description

I’m working with NVIDIA PeopleNet ResNet34 INT8 for object detection and need
help building a TensorRT engine from the encrypted .etlt model file. I’ve
attempted multiple approaches but encountered tooling compatibility issues.

Environment Details

Hardware:

  • GPU: NVIDIA A100 80GB PCIe (Compute Capability 8.0)
  • Platform: Azure NC96ads_A100_v4

Software:

  • OS: Ubuntu 22.04.5 LTS
  • Driver: 535.247.01
  • CUDA: 12.2
  • TensorRT: 10.13.0.35

Model Files:

  • resnet34_peoplenet_int8.etlt (86 MB)
  • resnet34_peoplenet_int8_calib.txt (9.4 KB)

What I’ve Tried

  1. TensorRT Python API - Cannot parse encrypted .etlt format
  2. Existing tao-converter - Requires TensorRT 8.x, I have 10.13
  3. DeepStream 7.0 Docker - Container environment complexities

What I Need

Option A (Preferred): Pre-built TensorRT engine file
Option B: Guidance on correct tao-converter version for TensorRT 10.13

Build Parameters

tao-converter
-k nvidia_tlt
-d 3,544,960
-o output_cov/Sigmoid,output_bbox/BiasAdd
-e resnet34_peoplenet_int8.engine
-t int8
-c resnet34_peoplenet_int8_calib.txt
-i nchw
-m 1
models/resnet34_peoplenet_int8.etlt

Questions

  1. Can someone provide a pre-built engine for A100 + TRT 10.13?
  2. What’s the correct tao-converter version for TensorRT 10.13?
  3. Should I convert .etlt to ONNX first?

Context

Production object detection pipeline for real-time video analytics. Pipeline is
ready - just blocked on TensorRT engine file.

Any help would be greatly appreciated! 🙏

Tags: deepstream, tensorrt, peoplenet, tao-toolkit, engine-build

2

Build command (if you have DeepStream/TAO):
tao-converter -k nvidia_tlt -d 3,544,960
-o output_cov/Sigmoid,output_bbox/BiasAdd
-e resnet34_peoplenet_int8.engine -t int8
-c resnet34_peoplenet_int8_calib.txt
models/resnet34_peoplenet_int8.etlt

Takes ~5 min with DeepStream. Can you help?

4

Is there any error? if so, could you share the error information?

5

duplicate with Request: Pre-built TensorRT Engine for PeopleNet ResNet34 (A100 + TRT 10.13)