最近在做Amlogic NPU想把流程记清楚。官方有开源仓库但对照README自己装一遍还是会踩到不少坑——尤其是国内网络、Git LFS、以及S905X5平台ID这件事。这篇就按我自己实际操作的顺序写走通这条链路yolov8n.pt → yolov8n.onnx → yolov8n_s905x5_w8a8.adla1. amlnn-toolkit是干什么的amlnn-toolkit是 Amlogic开源的一套转换工具负责导入.onnx/.tflite等模型量化本文用w8a8按目标芯片编译导出.adla可以简单理解为格式含义.ptPyTorch 权重.onnx中间交换格式.adla给指定 Amlogic NPU 用的可部署模型常见两种用法x86_64 Linux PC装amlnn_toolkit在主机上转换模型本文Amlogic Debian 板端装 edge toolkit在板子上跑相关仓库Amlogic-NN/amlnn-toolkitAmlogic-NN/amlnn-model-playground2. 安装 Miniconda / Miniforge工具依赖PyTorch、ONNX、TensorFlow一堆包建议独立Conda 环境别污染系统 Python。我这边实测机器是x86_64 Linux。如果本机还没有Conda先装Miniforge或 Miniconda 也行wgethttps://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-x86_64.shbashMiniforge3-Linux-x86_64.sh安装时按提示确认许可证、安装目录建议执行conda init。完成后source~/.bashrc conda--version如果不想每次自动进baseconda config--setauto_activate_basefalse然后创建 Python 3.10 环境公开仓库 README 和cp310wheel 都按 3.10 来conda create-namlnn_toolkit_py310python3.10-yconda activate amlnn_toolkit_py310 python--version3. 克隆amlnn-toolkit注意 Git LFS仓库里的 x86 wheel用 Git LFS 存的。如果只git clone没装 LFS后面你会发现.whl只有一百多字节——那是指针文件不是真正的包。3.1 安装Git LFSUbuntu / Debiansudoaptupdatesudoaptinstall-ygitgit-lfsgitlfsinstall3.2 克隆仓库gitclone https://github.com/Amlogic-NN/amlnn-toolkit.gitcdamlnn-toolkitgitlfs pull3.3 坑一clone GitHub较慢直连经常卡在几十 KB/s甚至超时。可以换镜像试试gitclone https://ghproxy.net/https://github.com/Amlogic-NN/amlnn-toolkit.gitcdamlnn-toolkitgitlfs pull注意镜像站点会变失效了就换别的代理 / 也可以本机下好再上传服务器或虚拟机。3.4 检查wheel是否真正下下来了ls-lhamlnn_toolkit/whl/linux_x86/正常情况类似amlnn_toolkit-1.0.0-cp310-cp310-linux_x86_64.whl 约 100MB amlnn_toolkit-1.0.0-cp311-...whl amlnn_toolkit-1.0.0-cp312-...whl requirements_x86_64.txt如果.whl只有134B左右百分之九十九是 LFS 没拉成功回去再执行gitlfsinstallgitlfs pull4. 安装Python依赖conda activate amlnn_toolkit_py310cdamlnn-toolkit官方 README写得是pipinstall-ramlnn_toolkit/whl/linux_x86/requirements_x86_64.txt pipinstallamlnn_toolkit/whl/linux_x86/amlnn_toolkit-1.0.0-cp310-cp310-linux_x86_64.whl实际在安装上大家不一定能一把梭成功。。4.1 添加镜像源网速慢的话添加阿里云镜像源能加速下载速度安装会需要一段时间。网好的可直接装。pipinstall-ramlnn_toolkit/whl/linux_x86/requirements_x86_64.txt-ihttp://mirrors.aliyun.com/pypi/simple pipinstallamlnn_toolkit/whl/linux_x86/amlnn_toolkit-1.0.0-cp310-cp310-linux_x86_64.whl-ihttp://mirrors.aliyun.com/pypi/simple4.2 坑二装完ai-edge-torch后import amlnn报 TF Lite so 缺符号原因大致是ai-edge-tensorflow和正式版tensorflow2.20.0打架。处理办法python-mpipinstall--force-reinstall --no-depstensorflow2.20.0-ihttps://pypi.tuna.tsinghua.edu.cn/simple4.3 安装 amlnn wheel依赖齐了之后再装 wheel。依赖已经手动搞定的话可以用--no-deps避免又回去解析 nightlypython-mpipinstall--no-deps amlnn_toolkit/whl/linux_x86/amlnn_toolkit-1.0.0-cp310-cp310-linux_x86_64.whl验证python-cfrom amlnn.api import AMLNN; print(amlnn-toolkit import OK)看到amlnn-toolkit import OK就说明这一关过了。5. 准备 YOLOv8n.pt→.onnx5.1 安装 Ultralytics别把 torch 冲掉如果已经有对应pt或者onnx模型文件可跳过此步骤toolkit 已经钉死torch2.9.1。如果直接pipinstallultralytics它可能顺手把 torch / CUDA 相关包装得乱七八糟。稳一点的话可以执行下面代码python-mpipinstall--no-deps ultralytics\-ihttps://pypi.tuna.tsinghua.edu.cn/simple python-mpipinstallmatplotlib pandas seaborn ultralytics-thop\py-cpuinfotorchvision0.24.1nvidia-ml-py polars\-ihttps://pypi.tuna.tsinghua.edu.cn/simple5.2 建工作目录mkdir-p~/amlnn_workspace/modelsmkdir-p~/amlnn_workspace/calibration/imagescd~/amlnn_workspace/models5.3yolov8n.pt下载Ultralytics 默认从 GitHub Releases 拉权重超时的话可以改用 HF 镜像curl-fL-oyolov8n.pt\https://hf-mirror.com/Ultralytics/YOLOv8/resolve/main/yolov8n.ptls-lhyolov8n.pt# 大约 6MB 左右5.4 导出 ONNX固定 640×640amlnn api输入格式支持ONNX、PyTorch、PyTorch pt2、TFLite这里使用ONNX输入。第一次转换建议先固定 shape少给后面编译添变量yoloexport\modelyolov8n.pt\formatonnx\imgsz640\batch1\dynamicFalse\simplifyTrue\opset13\nmsFalseAmlogic 官方 YOLOv8 脚本会切这三个内部输出Ultralytics 版本不同名字可能变/model.22/Concat_2_output_0 /model.22/Concat_1_output_0 /model.22/Concat_output_0我用 Ultralytics 8.4.x 导出后图里仍能看到这三个节点。建议用 Netron 再确认一眼。6. 准备量化校准图w8a8需要图片列表。把 jpg/png 放到~/amlnn_workspace/calibration/images/生成路径文件cd~/amlnn_workspacefind$(pwd)/calibration/images-typef\\(-iname*.jpg-o-iname*.jpeg-o-iname*.png\)\|sortcalibration/coco_subset.txtwc-lcalibration/coco_subset.txtheadcalibration/coco_subset.txt校准集对精度的影响以后再单独测。7. 导出 ADLAS905X5 / 005官方 playground 里有现成的 YOLOv8 转换脚本。可以整个 clonegitclone https://github.com/Amlogic-NN/amlnn-model-playground.git# 国内慢就同样用镜像也可以只把examples/yolov8/py/export_adla.py拷到本地例如import argparse import shutil from pathlib import Path import numpy as np from amlnn.api import AMLNN # Normalization constants used for quantization config MEAN np.array([0, 0, 0], dtypenp.float32) STD np.array([255, 255, 255], dtypenp.float32) def snapshot_adla_files(search_dir): return {path: path.stat().st_mtime for path in search_dir.rglob(*.adla)} def find_updated_adla_files(search_dir, known_files): current_files snapshot_adla_files(search_dir) updated_files [ path for path, mtime in current_files.items() if path not in known_files or mtime known_files[path] ] return sorted( updated_files, keylambda path: path.stat().st_mtime, reverseTrue, ) def main(): parser argparse.ArgumentParser(descriptionExport ONNX to ADLA) parser.add_argument(--onnx, requiredTrue, helpPath to ONNX model) parser.add_argument(--dataset-path, requiredTrue, helpPath to quant dataset) parser.add_argument(--target-platform, requiredTrue, helpPlatform ID, e.g. 001, 002, 003) parser.add_argument(--adla, default../model, helpOptional output .adla path) args parser.parse_args() search_dir Path.cwd() known_adla_files snapshot_adla_files(search_dir) if args.adla else {} amlnn AMLNN() # NOTE: These node names may be different depending on your model amlnn.load_onnx( modelargs.onnx, outputs[ /model.22/Concat_2_output_0, # -- Stride 32 (1x144x13x13 grid) /model.22/Concat_1_output_0, # -- Stride 16 (1x144x2x26 grid) /model.22/Concat_output_0 # -- Stride 8 (1x144x52x52 grid) ]) amlnn.config( normalization_mean[MEAN.tolist()], normalization_std[STD.tolist()], quantized_dtypew8a8, target_platformfPRODUCT_PID0XA{args.target_platform.zfill(3)}, ) amlnn.compile(datasetargs.dataset_path) amlnn.export_adla() if args.adla: new_adla_files find_updated_adla_files(search_dir, known_adla_files) if not new_adla_files: raise RuntimeError(export_adla did not create or update a .adla file) output_path Path(args.adla) output_path.parent.mkdir(parentsTrue, exist_okTrue) if new_adla_files[0].resolve() ! output_path.resolve(): shutil.copy2(new_adla_files[0], output_path) print(fsaved: {output_path}) if __name__ __main__: main()核心就是四步amlnn.load_onnx(...)amlnn.config(...)amlnn.compile(...)amlnn.export_adla()关键配置amlnn.config(normalization_mean[[0.0,0.0,0.0]],normalization_std[[255.0,255.0,255.0]],quantized_dtypew8a8,target_platformfPRODUCT_PID0XA{target_platform},)执行python ~/amlnn_workspace/scripts/export_adla.py\--onnx~/amlnn_workspace/models/yolov8n.onnx\--dataset-path ~/amlnn_workspace/calibration/coco_subset.txt\--target-platform 005\--adla~/amlnn_workspace/models/yolov8n_s905x5_w8a8.adla成功日志里应出现ls-lh~/amlnn_workspace/models/yolov8n_s905x5_w8a8.adla用 adlainfo 看一眼结果# 仓库路径 amlnn-toolkit/amlnn_toolkit/function_tool/02_adla_info_tool/ cd ~/amlnn-toolkit/amlnn_toolkit/function_tool/02_adla_info_tool chmod x adlainfo_release ./adlainfo_release -v ~/amlnn_workspace/models/yolov8n_s905x5_w8a8.adla ./adlainfo_release -io ~/amlnn_workspace/models/yolov8n_s905x5_w8a8.adla实测大致是Compile Version3.4.4Aml Hw Version4.0输入1×640×640×3Int8输出三路检测相关张量Int8参考Amlogic-NN/amlnn-toolkitAmlogic-NN/amlnn-model-playgroundYOLOv8 示例Ultralytics ExportONNX本文基于公开仓库个人实测整理仅供学习记录。工具版本和平台映射可能变化请以当下 README / release note 为准。欢迎转载请注明出处。
Amlogic NPU 上手（一）：安装amlnn-toolkit，YOLOv8n 导出ADLA
最近在做Amlogic NPU想把流程记清楚。官方有开源仓库但对照README自己装一遍还是会踩到不少坑——尤其是国内网络、Git LFS、以及S905X5平台ID这件事。这篇就按我自己实际操作的顺序写走通这条链路yolov8n.pt → yolov8n.onnx → yolov8n_s905x5_w8a8.adla1. amlnn-toolkit是干什么的amlnn-toolkit是 Amlogic开源的一套转换工具负责导入.onnx/.tflite等模型量化本文用w8a8按目标芯片编译导出.adla可以简单理解为格式含义.ptPyTorch 权重.onnx中间交换格式.adla给指定 Amlogic NPU 用的可部署模型常见两种用法x86_64 Linux PC装amlnn_toolkit在主机上转换模型本文Amlogic Debian 板端装 edge toolkit在板子上跑相关仓库Amlogic-NN/amlnn-toolkitAmlogic-NN/amlnn-model-playground2. 安装 Miniconda / Miniforge工具依赖PyTorch、ONNX、TensorFlow一堆包建议独立Conda 环境别污染系统 Python。我这边实测机器是x86_64 Linux。如果本机还没有Conda先装Miniforge或 Miniconda 也行wgethttps://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-x86_64.shbashMiniforge3-Linux-x86_64.sh安装时按提示确认许可证、安装目录建议执行conda init。完成后source~/.bashrc conda--version如果不想每次自动进baseconda config--setauto_activate_basefalse然后创建 Python 3.10 环境公开仓库 README 和cp310wheel 都按 3.10 来conda create-namlnn_toolkit_py310python3.10-yconda activate amlnn_toolkit_py310 python--version3. 克隆amlnn-toolkit注意 Git LFS仓库里的 x86 wheel用 Git LFS 存的。如果只git clone没装 LFS后面你会发现.whl只有一百多字节——那是指针文件不是真正的包。3.1 安装Git LFSUbuntu / Debiansudoaptupdatesudoaptinstall-ygitgit-lfsgitlfsinstall3.2 克隆仓库gitclone https://github.com/Amlogic-NN/amlnn-toolkit.gitcdamlnn-toolkitgitlfs pull3.3 坑一clone GitHub较慢直连经常卡在几十 KB/s甚至超时。可以换镜像试试gitclone https://ghproxy.net/https://github.com/Amlogic-NN/amlnn-toolkit.gitcdamlnn-toolkitgitlfs pull注意镜像站点会变失效了就换别的代理 / 也可以本机下好再上传服务器或虚拟机。3.4 检查wheel是否真正下下来了ls-lhamlnn_toolkit/whl/linux_x86/正常情况类似amlnn_toolkit-1.0.0-cp310-cp310-linux_x86_64.whl 约 100MB amlnn_toolkit-1.0.0-cp311-...whl amlnn_toolkit-1.0.0-cp312-...whl requirements_x86_64.txt如果.whl只有134B左右百分之九十九是 LFS 没拉成功回去再执行gitlfsinstallgitlfs pull4. 安装Python依赖conda activate amlnn_toolkit_py310cdamlnn-toolkit官方 README写得是pipinstall-ramlnn_toolkit/whl/linux_x86/requirements_x86_64.txt pipinstallamlnn_toolkit/whl/linux_x86/amlnn_toolkit-1.0.0-cp310-cp310-linux_x86_64.whl实际在安装上大家不一定能一把梭成功。。4.1 添加镜像源网速慢的话添加阿里云镜像源能加速下载速度安装会需要一段时间。网好的可直接装。pipinstall-ramlnn_toolkit/whl/linux_x86/requirements_x86_64.txt-ihttp://mirrors.aliyun.com/pypi/simple pipinstallamlnn_toolkit/whl/linux_x86/amlnn_toolkit-1.0.0-cp310-cp310-linux_x86_64.whl-ihttp://mirrors.aliyun.com/pypi/simple4.2 坑二装完ai-edge-torch后import amlnn报 TF Lite so 缺符号原因大致是ai-edge-tensorflow和正式版tensorflow2.20.0打架。处理办法python-mpipinstall--force-reinstall --no-depstensorflow2.20.0-ihttps://pypi.tuna.tsinghua.edu.cn/simple4.3 安装 amlnn wheel依赖齐了之后再装 wheel。依赖已经手动搞定的话可以用--no-deps避免又回去解析 nightlypython-mpipinstall--no-deps amlnn_toolkit/whl/linux_x86/amlnn_toolkit-1.0.0-cp310-cp310-linux_x86_64.whl验证python-cfrom amlnn.api import AMLNN; print(amlnn-toolkit import OK)看到amlnn-toolkit import OK就说明这一关过了。5. 准备 YOLOv8n.pt→.onnx5.1 安装 Ultralytics别把 torch 冲掉如果已经有对应pt或者onnx模型文件可跳过此步骤toolkit 已经钉死torch2.9.1。如果直接pipinstallultralytics它可能顺手把 torch / CUDA 相关包装得乱七八糟。稳一点的话可以执行下面代码python-mpipinstall--no-deps ultralytics\-ihttps://pypi.tuna.tsinghua.edu.cn/simple python-mpipinstallmatplotlib pandas seaborn ultralytics-thop\py-cpuinfotorchvision0.24.1nvidia-ml-py polars\-ihttps://pypi.tuna.tsinghua.edu.cn/simple5.2 建工作目录mkdir-p~/amlnn_workspace/modelsmkdir-p~/amlnn_workspace/calibration/imagescd~/amlnn_workspace/models5.3yolov8n.pt下载Ultralytics 默认从 GitHub Releases 拉权重超时的话可以改用 HF 镜像curl-fL-oyolov8n.pt\https://hf-mirror.com/Ultralytics/YOLOv8/resolve/main/yolov8n.ptls-lhyolov8n.pt# 大约 6MB 左右5.4 导出 ONNX固定 640×640amlnn api输入格式支持ONNX、PyTorch、PyTorch pt2、TFLite这里使用ONNX输入。第一次转换建议先固定 shape少给后面编译添变量yoloexport\modelyolov8n.pt\formatonnx\imgsz640\batch1\dynamicFalse\simplifyTrue\opset13\nmsFalseAmlogic 官方 YOLOv8 脚本会切这三个内部输出Ultralytics 版本不同名字可能变/model.22/Concat_2_output_0 /model.22/Concat_1_output_0 /model.22/Concat_output_0我用 Ultralytics 8.4.x 导出后图里仍能看到这三个节点。建议用 Netron 再确认一眼。6. 准备量化校准图w8a8需要图片列表。把 jpg/png 放到~/amlnn_workspace/calibration/images/生成路径文件cd~/amlnn_workspacefind$(pwd)/calibration/images-typef\\(-iname*.jpg-o-iname*.jpeg-o-iname*.png\)\|sortcalibration/coco_subset.txtwc-lcalibration/coco_subset.txtheadcalibration/coco_subset.txt校准集对精度的影响以后再单独测。7. 导出 ADLAS905X5 / 005官方 playground 里有现成的 YOLOv8 转换脚本。可以整个 clonegitclone https://github.com/Amlogic-NN/amlnn-model-playground.git# 国内慢就同样用镜像也可以只把examples/yolov8/py/export_adla.py拷到本地例如import argparse import shutil from pathlib import Path import numpy as np from amlnn.api import AMLNN # Normalization constants used for quantization config MEAN np.array([0, 0, 0], dtypenp.float32) STD np.array([255, 255, 255], dtypenp.float32) def snapshot_adla_files(search_dir): return {path: path.stat().st_mtime for path in search_dir.rglob(*.adla)} def find_updated_adla_files(search_dir, known_files): current_files snapshot_adla_files(search_dir) updated_files [ path for path, mtime in current_files.items() if path not in known_files or mtime known_files[path] ] return sorted( updated_files, keylambda path: path.stat().st_mtime, reverseTrue, ) def main(): parser argparse.ArgumentParser(descriptionExport ONNX to ADLA) parser.add_argument(--onnx, requiredTrue, helpPath to ONNX model) parser.add_argument(--dataset-path, requiredTrue, helpPath to quant dataset) parser.add_argument(--target-platform, requiredTrue, helpPlatform ID, e.g. 001, 002, 003) parser.add_argument(--adla, default../model, helpOptional output .adla path) args parser.parse_args() search_dir Path.cwd() known_adla_files snapshot_adla_files(search_dir) if args.adla else {} amlnn AMLNN() # NOTE: These node names may be different depending on your model amlnn.load_onnx( modelargs.onnx, outputs[ /model.22/Concat_2_output_0, # -- Stride 32 (1x144x13x13 grid) /model.22/Concat_1_output_0, # -- Stride 16 (1x144x2x26 grid) /model.22/Concat_output_0 # -- Stride 8 (1x144x52x52 grid) ]) amlnn.config( normalization_mean[MEAN.tolist()], normalization_std[STD.tolist()], quantized_dtypew8a8, target_platformfPRODUCT_PID0XA{args.target_platform.zfill(3)}, ) amlnn.compile(datasetargs.dataset_path) amlnn.export_adla() if args.adla: new_adla_files find_updated_adla_files(search_dir, known_adla_files) if not new_adla_files: raise RuntimeError(export_adla did not create or update a .adla file) output_path Path(args.adla) output_path.parent.mkdir(parentsTrue, exist_okTrue) if new_adla_files[0].resolve() ! output_path.resolve(): shutil.copy2(new_adla_files[0], output_path) print(fsaved: {output_path}) if __name__ __main__: main()核心就是四步amlnn.load_onnx(...)amlnn.config(...)amlnn.compile(...)amlnn.export_adla()关键配置amlnn.config(normalization_mean[[0.0,0.0,0.0]],normalization_std[[255.0,255.0,255.0]],quantized_dtypew8a8,target_platformfPRODUCT_PID0XA{target_platform},)执行python ~/amlnn_workspace/scripts/export_adla.py\--onnx~/amlnn_workspace/models/yolov8n.onnx\--dataset-path ~/amlnn_workspace/calibration/coco_subset.txt\--target-platform 005\--adla~/amlnn_workspace/models/yolov8n_s905x5_w8a8.adla成功日志里应出现ls-lh~/amlnn_workspace/models/yolov8n_s905x5_w8a8.adla用 adlainfo 看一眼结果# 仓库路径 amlnn-toolkit/amlnn_toolkit/function_tool/02_adla_info_tool/ cd ~/amlnn-toolkit/amlnn_toolkit/function_tool/02_adla_info_tool chmod x adlainfo_release ./adlainfo_release -v ~/amlnn_workspace/models/yolov8n_s905x5_w8a8.adla ./adlainfo_release -io ~/amlnn_workspace/models/yolov8n_s905x5_w8a8.adla实测大致是Compile Version3.4.4Aml Hw Version4.0输入1×640×640×3Int8输出三路检测相关张量Int8参考Amlogic-NN/amlnn-toolkitAmlogic-NN/amlnn-model-playgroundYOLOv8 示例Ultralytics ExportONNX本文基于公开仓库个人实测整理仅供学习记录。工具版本和平台映射可能变化请以当下 README / release note 为准。欢迎转载请注明出处。
