摘要本文详细记录了在 Linux 主机上使用 Qualcomm SNPE SDK 2.21 部署 VGG16 ONNX 模型的完整流程涵盖环境搭建、依赖安装、模型下载与转换、主机 CPU 推理验证、ARM64 交叉编译 C Sample、板端部署运行以及输出结果验证等九个步骤。文章还针对官方案例未预设 Python 环境的问题提供了多种虚拟环境配置方案帮助读者从零开始跑通 SNPE VGG 案例。一、相关文档入口这次 VGG 案例主要看这些文档主题文档SDK 环境docs/SNPE/html/general/setup.html教程资源准备docs/SNPE/html/general/tutorial_setup.htmlVGG ONNX 教程docs/SNPE/html/general/tutorial_onnx.htmlONNX 转 DLC 工具docs/SNPE/html/general/tools_model_conversion.html推理工具snpe-net-rundocs/SNPE/html/general/tools_execution.htmlC Sample 交叉编译docs/SNPE/html/general/cplus_plus_tutorial.htmlVGG 相关脚本在${SNPE_ROOT}/examples/Models/VGG/scripts/核心脚本setup_VGG.py create_VGG_raws.py create_file_list.py show_vgg_classifications.pyC 测试代码在${SNPE_ROOT}/examples/SNPE/NativeCpp/SampleCode/jni/main.cpp二、Linux 主机安装依赖SNPE 2.21 的check-python-dependency明确要求Python3.6或3.8必须在 virtualenv/conda 环境中运行推荐 Python 3.8如果你用uv推荐cd /path/to/snpe_2.21 uv python install 3.8 uv venv .venv-snpe --python 3.8 source .venv-snpe/bin/activate如果不用uvsudo apt update sudo apt install -y python3.8 python3.8-venv python3.8-dev python3-pip wget make build-essential cd /path/to/snpe_2.21 python3.8 -m venv .venv-snpe source .venv-snpe/bin/activate设置 SNPE 环境#XXX是snpe_2.21的本机路径自行替换正确路径后续执行模型转换后直接在onnx 模型路径下执行命令就行。export SNPE_ROOT/XXX/snpe_2.21 source ${SNPE_ROOT}/bin/envsetup.sh之前你遇到过Permission denied所以这里用python3直接执行检查脚本python3 ${SNPE_ROOT}/bin/check-python-dependencyVGG 最小依赖是python3 -m pip install numpy pillow scipy onnx但更推荐跑上面的check-python-dependency它会按 SDK 测试过的版本安装完整依赖。三、下载 VGG 模型和数据官方 VGG 脚本会下载三类资源vgg16.onnx synset.txt kitten.jpg来源在setup_VGG.py里写死https://s3.amazonaws.com/onnx-model-zoo/vgg/vgg16/vgg16.onnx https://s3.amazonaws.com/onnx-model-zoo/synset.txt https://s3.amazonaws.com/model-server/inputs/kitten.jpg执行cd ${SNPE_ROOT}/examples/Models/VGG python3 scripts/setup_VGG.py -a ./onnx -d脚本会自动完成1. 下载 vgg16.onnx、synset.txt、kitten.jpg 2. 生成 data/cropped/*.raw 3. 生成 data/cropped/raw_list.txt 4. 生成 dlc/vgg16.dlc注意跑测py需要在创建的uv环境下官方案例没有设定环境需要认为导入方法一激活虚拟环境推荐# 激活虚拟环境 source /home/XXX/My_Snpe_Project/snpe_2.21/.venv-snpe38/bin/activate 在虚拟环境中安装 onnx pip install onnx 验证 python -c import onnx; print(onnx.version)方法二查看当前用的是哪个 Pythonwhich python3 python3 --version 确认 onnx 安装位置 ls /home/XXX/.local/lib/python3.8/site-packages/ | grep onnx 临时把 user site-packages 加入 PYTHONPATH export PYTHONPATH/home/quectel/.local/lib/python3.8/site-packages:$PYTHONPATH python3 -c import onnx; print(onnx.version)将user site-packages 永久加入 PYTHONPATH方法方法一在虚拟环境中安装最推荐既然你已经有一个.venv-snpe38虚拟环境直接在虚拟环境里装onnx一劳永逸source /home/XXX/My_Snpe_Project/snpe_2.21/.venv-snpe38/bin/activate pip install onnx之后只要激活该虚拟环境onnx就能直接使用无需额外设置。方法二加到~/.bashrc全局生效如果你希望所有终端都自动包含 user site-packagesecho export PYTHONPATH/home/XXX/.local/lib/python3.8/site-packages:$PYTHONPATH ~/.bashrc source ~/.bashrc方法三修改 SNPE 的envsetup.sh仅在 source 时生效如果你希望只在 SNPE 环境下生效修改envsetup.shecho export PYTHONPATH/home/XXX/.local/lib/python3.8/site-packages:$PYTHONPATH /home/XXX/My_Snpe_Project/snpe_2.21/bin/envsetup.sh方法三直接在虚拟环境中安装如果已有虚拟环境# 用虚拟环境的 pip 安装 /home/XXX/My_Snpe_Project/snpe_2.21/.venv-snpe38/bin/pip install onnx 用虚拟环境的 python 运行 /home/XXX/My_Snpe_Project/snpe_2.21/.venv-snpe38/bin/python -c import onnx; print(onnx.version)如果网络不通也可以手动下载mkdir -p ${SNPE_ROOT}/examples/Models/VGG/onnx cd ${SNPE_ROOT}/examples/Models/VGG/onnx wget -N https://s3.amazonaws.com/onnx-model-zoo/vgg/vgg16/vgg16.onnx wget -N https://s3.amazonaws.com/onnx-model-zoo/synset.txt wget -N https://s3.amazonaws.com/model-server/inputs/kitten.jpg然后执行cd ${SNPE_ROOT}/examples/Models/VGG python3 scripts/create_VGG_raws.py -i onnx -d data/cropped python3 scripts/create_file_list.py -i data/cropped -o data/cropped/raw_list.txt -e *.raw python3 scripts/create_file_list.py -i data/cropped -o data/raw_list.txt -e *.raw -r cp onnx/synset.txt data/四、模型转换 ONNX → DLC如果不用setup_VGG.py自动转换可以手动执行cd ${SNPE_ROOT}/examples/Models/VGG mkdir -p dlc snpe-onnx-to-dlc --input_network onnx/vgg16.onnx --output_path dlc/vgg16.dlc或 在根目录下运行代码注意路径问题cd ${SNPE_ROOT}/examples/Models/VGG mkdir -p dlc python3 ./bin/x86_64-linux-clange/snpe-onnx-to-dlc --input_network ./examples/Models/VGG/onnx/vgg16.onnx --output_path ./examples/Models/VGG/dlc/vgg16.dlc检查 DLCsnpe-dlc-info -i dlc/vgg16.dlc或 在根目录下运行代码注意路径问题python3 ./bin/x86_64-linux-clang/snpe-dlc-info -i ./examples/Models/VGG/dlc/vgg16.dlc主机上先跑一次 CPU 推理验证 在根目录下运行代码注意路径问题cd ${SNPE_ROOT}/examples/Models/VGG/data/cropped snpe-net-run --input_list raw_list.txt --container ../../dlc/vgg16.dlc --output_dir ../../output注意snpe-net-run不是 Python 脚本是高通 SNPE 编译出来的二进制可执行文件你用python3去运行二进制文件必然报编码语法错误。补充常见前置操作赋予执行权限第一次运行必做# 替换成你自己的SNPE根目录 source ${SNPE_ROOT}/bin/envsetup.sh x86_64-linux-clang加载 SNPE 环境变量否则会报库找不到chmod x ./bin/x86_64-linux-clang/snpe-net-run查看分类结果cd ${SNPE_ROOT}/examples/Models/VGG python3 scripts/show_vgg_classifications.py -i data/cropped/raw_list.txt -o output -l data/synset.txt成功时会输出 top-5 分类例如 kitten 相关类别。举例 输入 python3 ./examples/Models/VGG/scripts/show_vgg_classifications.py -i ./examples/M odels/VGG/data/cropped/raw_list.txt -o ./examples/Models/VGG/output/ -l ./examples/Models/VGG/data/synset.txt输出 Classification results probability0.339871 ; classn02123045 tabby, tabby cat probability0.329623 ; classn02124075 Egyptian cat probability0.299326 ; classn02123159 tiger cat probability0.021479 ; classn02127052 lynx, catamount五、Linux 上交叉编译 ARM64 C SampleSNPE 已提供 Makefile${SNPE_ROOT}/examples/SNPE/NativeCpp/SampleCode/Makefile.aarch64-oe-linux-gcc11.2 ${SNPE_ROOT}/examples/SNPE/NativeCpp/SampleCode/Makefile.aarch64-ubuntu-gcc9.4如果你的板子是 Yocto / LE Linux推荐export SNPE_TARGET_ARCHaarch64-oe-linux-gcc11.2交叉工具链路径示例export AARCH64_LINUX_OE_GCC_112/path/to/qcom-esdk编译cd ${SNPE_ROOT}/examples/SNPE/NativeCpp/SampleCode make clean -f Makefile.aarch64-oe-linux-gcc11.2 make -f Makefile.aarch64-oe-linux-gcc11.2如果 Makefile 默认路径不匹配直接传CXXmake CXX/path/to/esdk/sysroots/x86_64-qtisdk-linux/usr/bin/aarch64-oe-linux/aarch64-oe-linux-g --sysroot/path/to/esdk/sysroots/armv8a-oe-linux \ -f Makefile.aarch64-oe-linux-gcc11.2产物${SNPE_ROOT}/examples/SNPE/NativeCpp/SampleCode/obj/local/aarch64-oe-linux-gcc11.2/snpe-sample如果你的目标板是 Ubuntu ARM64export SNPE_TARGET_ARCHaarch64-ubuntu-gcc9.4 export AARCH64_UBUNTU_GCC_94/path/to/aarch64-ubuntu-toolchain cd ${SNPE_ROOT}/examples/SNPE/NativeCpp/SampleCode make -f Makefile.aarch64-ubuntu-gcc9.4六、拷贝到 ARM Linux 设备假设板子 IP 是192.168.1.100用户名是rootexport SNPE_TARGET_ARCHaarch64-oe-linux-gcc11.2 export TARGET_IP192.168.1.100 export TARGET_USERroot export TARGET_DIR/data/local/tmp/snpe_vgg ssh ${TARGET_USER}${TARGET_IP} mkdir -p ${TARGET_DIR}/bin ${TARGET_DIR}/lib ${TARGET_DIR}/model ${TARGET_DIR}/data/cropped ${TARGET_DIR}/output拷贝 SNPE 运行库和工具scp ${SNPE_ROOT}/bin/${SNPE_TARGET_ARCH}/snpe-net-run \ ${TARGET_USER}${TARGET_IP}:${TARGET_DIR}/bin/ scp ${SNPE_ROOT}/lib/${SNPE_TARGET_ARCH}/*.so ${TARGET_USER}${TARGET_IP}:${TARGET_DIR}/lib/拷贝交叉编译的 samplescp ${SNPE_ROOT}/examples/SNPE/NativeCpp/SampleCode/obj/local/${SNPE_TARGET_ARCH}/snpe-sample \ ${TARGET_USER}${TARGET_IP}:${TARGET_DIR}/bin/拷贝 VGG 模型和输入scp ${SNPE_ROOT}/examples/Models/VGG/dlc/vgg16.dlc \ ${TARGET_USER}${TARGET_IP}:${TARGET_DIR}/model/ scp ${SNPE_ROOT}/examples/Models/VGG/data/cropped/*.raw ${TARGET_USER}${TARGET_IP}:${TARGET_DIR}/data/cropped/ scp ${SNPE_ROOT}/examples/Models/VGG/data/cropped/raw_list.txt ${TARGET_USER}${TARGET_IP}:${TARGET_DIR}/data/cropped/ scp ${SNPE_ROOT}/examples/Models/VGG/data/synset.txt ${TARGET_USER}${TARGET_IP}:${TARGET_DIR}/data/七、ARM 设备上运行 snpe-net-run登录板子ssh root192.168.1.100设置环境export TARGET_DIR/data/local/tmp/snpe_vgg export LD_LIBRARY_PATH${TARGET_DIR}/lib:${LD_LIBRARY_PATH} export PATH${TARGET_DIR}/bin:${PATH}测试 SNPE 工具snpe-net-run -h运行 VGGcd ${TARGET_DIR} snpe-net-run --input_list data/cropped/raw_list.txt --container model/vgg16.dlc --output_dir output输出文件一般是${TARGET_DIR}/output/Result_0/vgg0_dense2_fwd.raw八、ARM 设备上运行交叉编译 samplecd ${TARGET_DIR} snpe-sample -b ITENSOR -d model/vgg16.dlc -i data/cropped/raw_list.txt -o output_sample -r cpu参数说明-b ITENSOR 使用 ITensor 输入输出 -d vgg16.dlc DLC 模型 -i raw_list.txt 输入 raw 列表 -o output_sample 输出目录 -r cpu CPU runtime如果要试 GPU/DSPsnpe-sample -b ITENSOR -d model/vgg16.dlc -i data/cropped/raw_list.txt -o output_gpu -r gpu snpe-sample -b ITENSOR -d model/vgg16.dlc -i data/cropped/raw_list.txt -o output_dsp -r dspDSP 还需要额外推送 Hexagon 库并设置ADSP_LIBRARY_PATH先跑 CPU 最稳。九、测试代码拉回输出并验证 top-5在主机上拉回板端输出mkdir -p ${SNPE_ROOT}/examples/Models/VGG/arm_output scp -r root192.168.1.100:/data/local/tmp/snpe_vgg/output ${SNPE_ROOT}/examples/Models/VGG/arm_output/用官方测试脚本验证cd ${SNPE_ROOT}/examples/Models/VGG python3 scripts/show_vgg_classifications.py -i data/cropped/raw_list.txt -o arm_output/output -l data/synset.txt如果你想要一个更小的独立测试脚本可保存为verify_vgg_output.pyimport argparse import numpy as np from scipy.special import softmax parser argparse.ArgumentParser() parser.add_argument(--raw, requiredTrue) parser.add_argument(--labels, requiredTrue) args parser.parse_args() logits np.fromfile(args.raw, dtypenp.float32) if logits.size ! 1000: raise RuntimeError(fExpected 1000 outputs, got {logits.size}) with open(args.labels, r) as f: labels [x.strip() for x in f] probs softmax(logits) top5 np.argsort(probs)[::-1][:5] for i in top5: print(f{probs[i]:.6f} {labels[i]})运行python3 verify_vgg_output.py \ --raw ${SNPE_ROOT}/examples/Models/VGG/arm_output/output/Result_0/vgg0_dense2_fwd.raw \ --labels ${SNPE_ROOT}/examples/Models/VGG/data/synset.txt最短跑通路线source .venv-snpe/bin/activate export SNPE_ROOT/path/to/snpe_2.21 source ${SNPE_ROOT}/bin/envsetup.sh python3 ${SNPE_ROOT}/bin/check-python-dependency cd ${SNPE_ROOT}/examples/Models/VGG python3 scripts/setup_VGG.py -a ./onnx -d cd data/cropped snpe-net-run --input_list raw_list.txt --container ../../dlc/vgg16.dlc --output_dir ../../output cd ${SNPE_ROOT}/examples/Models/VGG python3 scripts/show_vgg_classifications.py -i data/cropped/raw_list.txt -o output -l data/synset.txt