PeopleNet model(.etlt) fails to work with JetPack4.6

nakayonum · July 11, 2025, 4:45am

Please provide the following information when requesting support.

• Hardware (T4/V100/Xavier/Nano/etc) Jetson Nano
• Network Type (Detectnet_v2/Faster_rcnn/Yolo_v4/LPRnet/Mask_rcnn/Classification/etc)
• TLT Version (Please run “tlt info --verbose” and share “docker_tag” here)
• Training spec file(If have, please share here)
• How to reproduce the issue ? (This is for errors. Please share the command line and the detailed log here.)

Hi,

I have JetPack4.6 on a Jetson Nano and want to change the model from resnet10.caffemodel to PeopleNet model pruned_quantized_v2.3.2 based on
[deepstream_test_3.py](https://github.com/NVIDIA-AI-IOT/deepstream_python_apps/tree/v1.1.1/apps/deepstream -test3)

I did the following steps, but could not find the object.
Please advise.

1.convert resnet34_peoplenet_pruned_int8.etlt in pruned_quantized_v2.3.2 to save.engine
1.1 Download TAO toolkit

wget -O jp46-20210820t231431z-001.zip https://developer.nvidia.com/jp46-20210820t231431z-001zip
mkdir jp46- 20210820t231431z-001
unzip -d jp46-20210820t231431z-001 jp46-20210820t231431z-001.zip

1.2- Download the model

Download and unzip https://catalog.ngc.nvidia.com/orgs/nvidia/teams/tao/models/peoplenet/versionからpruned_quantized_v2.3.2.zip

1.3 Generate engine file

desktop:~$ . /jp46-20210820t231431z-001/tao-converter-jp46-trt8.0.1.6/tao-converter -k tlt_encode -d 3,544,960 /root/models/peoplenet/peoplenet_ pruned_quantized_v2.3.2/resnet34_peoplenet_pruned_int8.etlt
desktop:~$
[INFO] [MemUsageChange] Init CUDA: CPU +203, GPU +0, now: CPU 221, GPU 2805 (MiB)
[INFO] [MemUsageSnapshot] Builder begin: CPU 239 MiB, GPU 2836 MiB
[INFO] ---------- Layers Running on DLA ----------
[INFO] ---------- Layers Running on GPU ----------
[INFO] [GpuLayer] conv1/convolution + activation_1/Relu6
[INFO] [GpuLayer] block_1a_conv_1/convolution + block_1a_relu_1/Relu6
[INFO] [GpuLayer] block_1a_conv_shortcut/convolution
[INFO] [GpuLayer] block_1a_conv_2/convolution + add_1/add + block_1a_relu/Relu6
[INFO] [GpuLayer] block_1b_conv_1/convolution + block_1b_relu_1/Relu6
[INFO] [GpuLayer] block_1b_conv_shortcut/convolution
[INFO] [GpuLayer] block_1b_conv_2/convolution + add_2/add + block_1b_relu/Relu6
[INFO] [GpuLayer] block_1c_conv_1/convolution + block_1c_relu_1/Relu6
[INFO] [GpuLayer] block_1c_conv_shortcut/convolution
[INFO] [GpuLayer] block_1c_conv_2/convolution + add_3/add + block_1c_relu/Relu6
[INFO] [GpuLayer] block_2a_conv_1/convolution + block_2a_relu_1/Relu6
[INFO] [GpuLayer] block_2a_conv_shortcut/convolution
[INFO] [GpuLayer] block_2a_conv_2/convolution + add_4/add + block_2a_relu/Relu6
[INFO] [GpuLayer] block_2b_conv_1/convolution + block_2b_relu_1/Relu6
[INFO] [GpuLayer] block_2b_conv_shortcut/convolution
[INFO] [GpuLayer] block_2b_conv_2/convolution + add_5/add + block_2b_relu/Relu6
[INFO] [GpuLayer] block_2c_conv_1/convolution + block_2c_relu_1/Relu6
[INFO] [GpuLayer] block_2c_conv_shortcut/convolution
[INFO] [GpuLayer] block_2c_conv_2/convolution + add_6/add + block_2c_relu/Relu6
[INFO] [GpuLayer] block_2d_conv_1/convolution + block_2d_relu_1/Relu6
[INFO] [GpuLayer] block_2d_conv_shortcut/convolution
[INFO] [GpuLayer] block_2d_conv_2/convolution + add_7/add + block_2d_relu/Relu6
[INFO] [GpuLayer] block_3a_conv_1/convolution + block_3a_relu_1/Relu6
[INFO] [GpuLayer] block_3a_conv_shortcut/convolution
[INFO] [GpuLayer] block_3a_conv_2/convolution + add_8/add + block_3a_relu/Relu6
[INFO] [GpuLayer] block_3b_conv_1/convolution + block_3b_relu_1/Relu6
[INFO] [GpuLayer] block_3b_conv_shortcut/convolution
[INFO] [GpuLayer] block_3b_conv_2/convolution + add_9/add + block_3b_relu/Relu6
[INFO] [GpuLayer] block_3c_conv_1/convolution + block_3c_relu_1/Relu6
[INFO] [GpuLayer] block_3c_conv_shortcut/convolution
[INFO] [GpuLayer] block_3c_conv_2/convolution + add_10/add + block_3c_relu/Relu6
[INFO] [GpuLayer] block_3d_conv_1/convolution + block_3d_relu_1/Relu6
[INFO] [GpuLayer] block_3d_conv_shortcut/convolution
[INFO] [GpuLayer] block_3d_conv_2/convolution + add_11/add + block_3d_relu/Relu6
[INFO] [GpuLayer] block_3e_conv_1/convolution + block_3e_relu_1/Relu6
[INFO] [GpuLayer] block_3e_conv_shortcut/convolution
[INFO] [GpuLayer] block_3e_conv_2/convolution + add_12/add + block_3e_relu/Relu6
[INFO] [GpuLayer] block_3f_conv_1/convolution + block_3f_relu_1/Relu6
[INFO] [GpuLayer] block_3f_conv_shortcut/convolution
[INFO] [GpuLayer] block_3f_conv_2/convolution + add_13/add + block_3f_relu/Relu6
[INFO] [GpuLayer] block_4a_conv_1/convolution + block_4a_relu_1/Relu6
[INFO] [GpuLayer] block_4a_conv_shortcut/convolution
[INFO] [GpuLayer] block_4a_conv_2/convolution + add_14/add + block_4a_relu/Relu6
[INFO] [GpuLayer] block_4b_conv_1/convolution + block_4b_relu_1/Relu6
[INFO] [GpuLayer] block_4b_conv_shortcut/convolution
[INFO] [GpuLayer] block_4b_conv_2/convolution + add_15/add + block_4b_relu/Relu6
[INFO] [GpuLayer] block_4c_conv_1/convolution + block_4c_relu_1/Relu6
[INFO] [GpuLayer] block_4c_conv_shortcut/convolution
[INFO] [GpuLayer] block_4c_conv_2/convolution + add_16/add + block_4c_relu/Relu6
[INFO] [GpuLayer] output_bbox/convolution
[INFO] [GpuLayer] output_cov/convolution
[INFO] [GpuLayer] PWN(output_cov/Sigmoid)
[INFO] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +158, GPU +154, now: CPU 406, GPU 2990 (MiB)
[INFO] [MemUsageChange] Init cuDNN: CPU +241, GPU +241, now: CPU 647, GPU 3231 (MiB)
[WARNING] Detected invalid timing cache, setup a local cache instead
[INFO] Some tactics do not have sufficient workspace memory to run. Increasing workspace size may increase performance, please check verbose output.
[INFO] Detected 1 inputs and 2 output network tensors.
[INFO] Total Host Persistent Memory: 78336
[INFO] Total Device Persistent Memory: 26554368
[INFO] Total Scratch Memory: 0
[INFO] [MemUsageStats] Peak memory usage of TRT CPU/GPU memory allocators: CPU 9 MiB, GPU 1239 MiB
[INFO] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +4, now: CPU 893, GPU 2886 (MiB)
[INFO] [MemUsageChange] Init cuDNN: CPU +0, GPU +2, now: CPU 893, GPU 2888 (MiB)
[INFO] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +0, now: CPU 893, GPU 2888 (MiB)
[INFO] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +0, now: CPU 892, GPU 2888 (MiB)
[INFO] [MemUsageSnapshot] Builder end: CPU 892 MiB, GPU 2888 MiB

save “save.engine”, “labels.txt” in “pruned_quantized_v2.3.2” and “resnet34_peoplenet_pruned_int8.txt” in “pruned_quantized_v2.3.2” to /root/models Save it in /root/models 2.
change the contents of dstest3_pgie_config.txt to the following:

[property]
gpu-id=0
net-scale-factor=1.0
model-engine-file=/root/models/peoplenet_pruned_quantized_v2.3.2/resnet34_peoplenet_pruned_int8.engine
labelfile-path=/root/models/peoplenet_pruned_quantized_v2.3.2/labels.txt
int8-calib-file=/root/models/peoplenet_pruned_quantized_v2.3.2/resnet34_peoplenet_int8.txt
batch-size=1
network-mode=1
num-detected-classes=3
interval=0
gie-unique-id=1
output-blob-names=output_bbox/Bias;output_cov/Sigmoid
cluster-mode=1
model-color-format=0
[class-attrs-all]
pre-cluster-threshold=0.5
topk=20
nms-iou-threshold=0.5

Run deepstream_test_3.py to show a person on the connected camera, but no object is detected

The print on line 101 detects objects, but does not print “Objects=1”.
I have not modified deepstream_test_3.py for peoplenet, but expect “Objects=” to be “1”

print(“Frame Number=”, frame_number, “Number of Objects=”,num_ rects, “Vehicle_count=”,obj_counter[PGIE_CLASS_ID_VEHICLE], “Person_count=”,obj_counter[PGIE_CLASS_ID_PERSON])

junshengy · July 11, 2025, 9:09am

This is a very legacy version. For pyds-1.1.1 to support peoplenet, you need to port the model and code yourself. It cannot run pruned_quantized_v2.3.2 directly.

You can refer to the guidance in deepstream_reference_apps(tag: DS_6.0). This peoplenet model has been verified on jp-4.6.

github.com/NVIDIA-AI-IOT/deepstream_reference_apps

deepstream_app_tao_configs/download_models.sh

DS_6.0


      
          -O trafficcamnet_pruned_v1.0.zip && \
          unzip trafficcamnet_pruned_v1.0.zip -d ../../models/tao_pretrained_models/trafficcamnet && \
          rm trafficcamnet_pruned_v1.0.zip
          mkdir -p ../../models/tao_pretrained_models/facedetectir && \
          wget --content-disposition https://api.ngc.nvidia.com/v2/models/nvidia/tao/facedetectir/versions/pruned_v1.0/zip \
          -O facedetectir_pruned_v1.0.zip && \
          unzip facedetectir_pruned_v1.0.zip -d ../../models/tao_pretrained_models/facedetectir && \
          rm facedetectir_pruned_v1.0.zip
          
          echo "==================================================================="
          echo "begin download models for peopleNet "
          echo "==================================================================="
          mkdir -p ../../models/tao_pretrained_models/peopleNet/V2.1
          wget --content-disposition https://api.ngc.nvidia.com/v2/models/nvidia/tao/peoplenet/versions/pruned_quantized_v2.1/zip \
          -O peoplenet_pruned_int8_v2.1.zip && \
          unzip peoplenet_pruned_int8_v2.1.zip -d ../../models/tao_pretrained_models/peopleNet/V2.1 && \
          rm peoplenet_pruned_int8_v2.1.zip
          
          echo "==================================================================="
          echo "begin download models for peopleSegNet "
          echo "==================================================================="

nakayonum · July 14, 2025, 1:52am

Hi,
Thanks for the advice
I was able to get a detection with the “pruned_quantized_v2.1” you gave me.

However, I would like to run other Peoplenetmodels to check the accuracy of detection, etc.

Is there any
model of PeopleNet Model that can be run in .etlt format?
If so, please let me know the nvinfer configuration file (config_infer_) and an example of using tao-converter.

junshengy · July 14, 2025, 6:11am

v2.1 is already a model in etlt format, and this repository above also contains configuration file

github.com/NVIDIA-AI-IOT/deepstream_reference_apps

deepstream_app_tao_configs/config_infer_primary_peoplenet.txt

DS_6.0

################################################################################
# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
#
# Permission is hereby granted, free of charge, to any person obtaining a
# copy of this software and associated documentation files (the "Software"),
# to deal in the Software without restriction, including without limitation
# the rights to use, copy, modify, merge, publish, distribute, sublicense,
# and/or sell copies of the Software, and to permit persons to whom the
# Software is furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
# DEALINGS IN THE SOFTWARE.

This file has been truncated. show original

nakayonum · July 14, 2025, 8:44am

The question was not accurate and seems to have caused misunderstanding.
I will ask the question again.

I want to run the following Version .etlt files at
[PeopleNet | NVIDIA NGC] They are

deployable_quantized_v2.5
deployable_quantized_v2.5.2
deployable_quantized_v2.6
deployable_quantized_v2.6.1
pruned_quantized_v2.1
pruned_quantized_v2.1.1
pruned_quantized_v2.3
pruned_quantized_v2.3.2 ← I understand that creating .engine will not work due to legacy versions.
pruned_v1.0
pruned_v2.0
pruned_v2.1
pruned_v2.3
pruned_quantized_v2.3.2

Please let me know the “nvinfer configuration file (config_infer_)” and “how to convert .etlt to .engine (example of using tao-converter)” needed to get it working.
Do you mean that the “nvinfer configuration file (config_infer_)” for all versions is “deepstream_app_tao_configs/config_infer_primary_peoplenet.txt”
?

junshengy · July 14, 2025, 9:40am

We cannot guarantee that other versions will run correctly with Jetpack 4.6. In other words, the peoplenet version is related to the jetpack version. Because tensorrt is part of jetpack.
If you want better results, you can finetune pruned_quantized_v2.1.

For deepstream 6.0, nvinfer will automatically convert the model from etlt to engine.

For tao-converter, refer to this command line.

./tao-converter -k tlt_encode -t int8 -c resnet34_peoplenet_pruned_int8_gpu.txt  -b 1 -d 3,544,960 -e resnet34_peoplenet_int8.etlt_b1_gpu0_int8.engine resnet34_peoplenet_pruned_int8.etlt

nakayonum · July 15, 2025, 2:31am

Thanks ,

We cannot guarantee that other versions will run correctly with Jetpack 4.6.
In other words, the peoplenet version is related to the jetpack version.
Because tensorrt is part of jetpack.

Do you have a list of the versions of the people model and the corresponding versions of jetapck?

junshengy · July 15, 2025, 10:04am

The tags in the deepstream_reference_apps repository correspond to different versions of deepstream, and the download_models.sh files with different tags show compatible versions.

nakayonum · July 16, 2025, 12:07am

Understood. Thank you very much.

system · July 30, 2025, 12:08am

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.

Topic		Replies	Views
Use jetson-inference TensorRT to infer from custom models like PeopleNet Jetson Nano tensorrt , jetson-inference	8	858	October 18, 2021
transfert learning toolkit-> export model TAO Toolkit	11	3612	October 12, 2021
Python App Cutom Model on the Jetson Nano TAO Toolkit	10	1188	October 12, 2021
Accelerating Peoplnet with tlt for jetson nano TAO Toolkit	19	2464	October 12, 2021
Unable to use custom-trained peoplenet model TAO Toolkit	6	892	April 20, 2022
Run peoplenet on python virtual environment without jetson TAO Toolkit tensorrt	29	1269	January 20, 2024
Fpenet retraining output file onnx but deepstream is using tlt TAO Toolkit	22	933	October 17, 2023
Peoplenet model error on Jetson Orin Nano Jetson Orin Nano jetson-inference	6	505	February 29, 2024
Not able to deploy .etlt file in deepstream test app 1 TAO Toolkit	12	1846	October 12, 2021
Launching customized .etlt model on jetson DeepStream SDK	3	402	January 25, 2022

PeopleNet model(.etlt) fails to work with JetPack4.6

Related topics