NPU Quick Validation

This document provides an out-of-the-box resnet50 object recognition model inference example. This example performs hardware-accelerated inference using Qualcomm® Hexagon™ Processor (NPU) on Radxa Dragon Ubuntu system.

tip

Before performing NPU Quick Verification, please refer to Enable NPU on Board to make sure the NPU is enabled.

Download Example

Ubuntu 24.04

Device

pip3 install modelscope
modelscope download --model radxa/resnet50_qairt --local ./resnet50_qairt

Ubuntu 26.04

Device

pip3 install modelscope
modelscope download --model radxa/resnet50_qairt --local-dir ./resnet50_qairt

Run Example

Please import environment variables according to SoC

QCS6490

Device

export PRODUCT_SOC=6490

SC8280XP

Device

export PRODUCT_SOC=8280

QCS9075

Device

export PRODUCT_SOC=9075

Execute model inference

Device

cd resnet50_qairt/${PRODUCT_SOC}
chmod +x qnn-net-run
./qnn-net-run --backend ./libQnnHtp.so --retrieve_context ./resnet50_aimet_quantized_${PRODUCT_SOC}.bin --input_list ./test_list.txt --output_dir output_bin

Verify Example

You can use python script for result verification

Device

cd ../scripts
python3 show_resnet50_classifications.py --input_list ../${PRODUCT_SOC}/test_list.txt -o ../${PRODUCT_SOC}/output_bin/ --labels_file ../data/imagenet_classes.txt

$ python3 show_resnet50_classifications.py --input_list ../${PRODUCT_SOC}/test_list.txt -o ../${PRODUCT_SOC}/output_bin/ --labels_file ../data/imagenet_classes.txt
Classification results
../data/test/crop/ILSVRC2012_val_00003441.raw 21.476509 402 acoustic guitar
../data/test/crop/ILSVRC2012_val_00008465.raw 22.651005 927 trifle
../data/test/crop/ILSVRC2012_val_00010218.raw 12.248322 281 tabby
../data/test/crop/ILSVRC2012_val_00044076.raw 18.456375 376 proboscis monkey

By comparing the printed results with the test image content, you can confirm that the output results of the resnet50 model ported to Qualcomm® NPU are correct.

resnet50 input images