当前位置：首页 > news >正文

Qualcomm SNPE（Neural Processing SDK）集成到 OpenWRT + QCS6490 的完整配置后续

news 2025/10/25 11:55:03

基于前面的内容本期继续接着开发过程来说，下面给出 三套完整可直接使用的源码，全部基于 OpenWRT + QCS6490 + SNPE 环境：

自动化模型转换脚本（主机端，ONNX → DLC + 量化 + 校验）
LuCI 推理控制面板（Web 上传模型、选择运行时、实时推理）
OTA 模型更新服务（设备端自动下载、校验、替换、通知）

1. 自动化模型转换脚本（主机端）

文件：convert_model.sh
路径：~/snpe-tools/convert_model.sh
功能：一键 ONNX → DLC → 量化 → 校验 → 打包 OTA 包

#!/bin/bash
# convert_model.sh
# 用法: ./convert_model.sh yolov8n.onnx yolov8nset -eMODEL_ONNX="$1"
MODEL_NAME="$2"
INPUT_SIZE="${3:-640}"  # 默认 640x640
QUANTIZE="${4:-true}"   # 是否量化
SNPE_ROOT="$HOME/snpe-sdk"
OUTPUT_DIR="./ota_models"# 1. 检查环境
if [ ! -d "$SNPE_ROOT" ]; thenecho "SNPE_ROOT not found: $SNPE_ROOT"exit 1
fisource "$SNPE_ROOT/bin/envsetup.sh"mkdir -p "$OUTPUT_DIR"DLC_FILE="${OUTPUT_DIR}/${MODEL_NAME}.dlc"
DLC_QUANT="${OUTPUT_DIR}/${MODEL_NAME}_quant.dlc"
INPUT_LIST="${OUTPUT_DIR}/input_list.txt"
OTA_PKG="${OUTPUT_DIR}/${MODEL_NAME}_ota.tar.gz"# 2. 生成输入列表（随机图像）
echo "Generating input list..."
mkdir -p images
for i in {1..5}; doconvert -size ${INPUT_SIZE}x${INPUT_SIZE} xc:gray images/calib_$i.jpgecho "images/calib_$i.jpg" >> "$INPUT_LIST"
done# 3. ONNX → DLC
echo "Converting ONNX to DLC..."
snpe-onnx-to-dlc \--input_model "$MODEL_ONNX" \--output_model "$DLC_FILE" \--input_dim input_0 1,3,${INPUT_SIZE},${INPUT_SIZE}# 4. 量化（可选）
if [ "$QUANTIZE" = "true" ]; thenecho "Quantizing to INT8..."snpe-dlc-quantize \--input_dlc "$DLC_FILE" \--input_list "$INPUT_LIST" \--output_dlc "$DLC_QUANT" \--use_enhanced_quantizer \--use_adjusted_weights_quantizerFINAL_DLC="$DLC_QUANT"
elseFINAL_DLC="$DLC_FILE"
fi# 5. 校验模型
echo "Validating model..."
snpe-dlc-info --input_dlc "$FINAL_DLC" | head -20# 6. 打包 OTA 包（含校验码）
echo "Packaging OTA update..."
cat > "${OUTPUT_DIR}/model.json" << EOF
{"name": "$MODEL_NAME","version": "$(date +%Y%m%d%H%M)","size": $(stat -c%s "$FINAL_D cellular"),"sha256": "$(sha256sum "$FINAL_DLC" | cut -d' ' -f1)","input_size": ${INPUT_SIZE},"quantized": $QUANTIZE
}
EOFtar -czf "$OTA_PKG" -C "$OUTPUT_DIR" \"$(basename $FINAL_DLC)" model.jsonecho "OTA package ready: $OTA_PKG"
echo "Upload to: http://your-server.com/ota/${MODEL_NAME}_ota.tar.gz"

使用示例：

./convert_model.sh yolov8n.onnx yolov8n 640 true

2. LuCI 推理控制面板（Web 界面）

路径：package/luci-app-snpe/

目录结构

luci-app-snpe/
├── root/
│   └── usr/share/luci/menu.d/snpe.json
├── htdocs/
│   └── luci-static/resources/view/snpe/
│       ├── dashboard.htm
│       └── inference.js
└── src/└── Makefile

`src/Makefile`

include $(TOPDIR)/rules.mkPKG_NAME:=luci-app-snpe
PKG_VERSION:=1.0
PKG_RELEASE:=1include $(INCLUDE_DIR)/package.mkdefine Package/luci-app-snpeSECTION:=luciCATEGORY:=LuCISUBMENU:=3. ApplicationsTITLE:=SNPE AI Inference PanelDEPENDS:=+snpe +python3 +python3-opencv +luci-base
endefdefine Build/Compile
endefdefine Package/luci-app-snpe/install$(INSTALL_DIR) $(1)/usr/share/luci/menu.d$(CP) ./root/usr/share/luci/menu.d/snpe.json $(1)/usr/share/luci/menu.d/$(INSTALL_DIR) $(1)/usr/share/luci/static/resources/view/snpe$(CP) ./htdocs/luci-static/resources/view/snpe/* $(1)/usr/share/luci/static/resources/view/snpe/
endef$(eval $(call BuildPackage,luci-app-snpe))

`root/usr/share/luci/menu.d/snpe.json`

{"admin/ai": {"snpe": {"title": "SNPE Inference","order": 10,"action": {"type": "view","path": "snpe/dashboard"}}}
}

`htdocs/luci-static/resources/view/snpe/dashboard.htm`

<%+header%>
<h2>SNPE AI Inference Panel</h2><div class="cbi-section"><div class="cbi-section-node"><label class="cbi-value-title">Model</label><div class="cbi-value-field"><select id="model_select" class="cbi-input-select"></select><button onclick="uploadModel()">Upload New</button><input type="file" id="model_file" style="display:none" accept=".dlc"></div></div><div class="cbi-section-node"><label>Runtime</label><select id="runtime"><option value="DSP">DSP (Hexagon)</option><option value="HTP">HTP (NPU)</option></select></div><div class="cbi-section-node