国产深度学习框架paddlepaddle与国产世界首款AI芯片，未能形成有效的通路，非常让人遗憾。在海思缺货的今天，地平线旭日系列是很好的替代品，此项目讲述如何实现飞桨模型在地平线开发板的部署！
☞☞☞AI 智能聊天, 问答助手, AI 智能搜索, 免费无限量使用 DeepSeek R1 模型☜☜☜
大家好，我是海盗旗，资深CV工程师（ctrl c + ctrl v），第一次使用国产框架与国产AI芯片（地平线旭日X3）之间实现了全链路开发。
本次教程，使用飞桨高阶API训练mobilenet模型，三分钟即可搭建一个网络，进行训练。
部署使用地平线的天工开物工具链，模型快速转化上板，极短的时间内可演示demo。
欢迎大家一起使用国产化平台进行AI开发与部署！
整体开发流程，来个图感受下
onnx模型是paddle模型与开发板推理部署之间的桥梁，模型的转化是其中非常重要的一环。
本次教程主要讲述流程串通，直接使用飞浆的高阶api训练模型,可以节省大量的开发时间
一、模型训练：
飞桨高阶api对基础aoi进行了封装，模型调用与数据预处理几行代码就可以搞定，真的是灰常方便，建议大家多多使用。
本次教程使用mobilenetv2进行分类，数据预处理的时候，模型标签是从1-102，但是深度学习标签都是从0开始，建议百度的工程师修改一下此处的标签。 本次教程使用mobilenetv2进行分类，数据预处理的时候，模型标签是从1-102，但是深度学习标签都是从0开始。
修改的方法需要把标签数直接减去1，这样标签数据对齐到0-101，可以直接训练，否则会报错，标签索引超出！
In [ ]
#只要20行代码，深度学习模型带回家，高阶api，用起来吧import paddle
from paddle.vision.datasets import Flowersfrom paddle.vision.models import MobileNetV2,mobilenet_v2from paddle.vision.transforms import ToTensor,Compose, Resize, ColorJitterprint('===============================start train')
transform = Compose([Resize(size=[228,228]),ToTensor()])#预处理train_dataset = Flowers(mode='train',transform=transform)#标签转换为0-101，不改的话，下面的标签需要改成103train_dataset.labels = train_dataset.labels-1 #花有102种类别#创建模型mobilenetv2 = MobileNetV2(num_classes=102)# paddle.summary(mobilenetv2,(1,3,228,228)) 测试模型model = paddle.Model(mobilenetv2)#模型封装model.load('./Model/7')#一天只有十个小时，太难了，只能分段训练、、、model.prepare(paddle.optimizer.Adam(parameters=model.parameters()),
paddle.nn.CrossEntropyLoss(),
paddle.metric.Accuracy())
model.fit(train_dataset,
epochs=200,
batch_size=64,
save_dir='Model/cpu',
save_freq=1,
verbose=1)print('===============================finish train')===============================start train
Cache file /home/aistudio/.cache/paddle/dataset/flowers/102flowers.tgz not found, downloading http://paddlemodels.bj.bcebos.com/flowers/102flowers.tgz Begin to download Download finished Cache file /home/aistudio/.cache/paddle/dataset/flowers/imagelabels.mat not found, downloading http://paddlemodels.bj.bcebos.com/flowers/imagelabels.mat Begin to download . Download finished Cache file /home/aistudio/.cache/paddle/dataset/flowers/setid.mat not found, downloading http://paddlemodels.bj.bcebos.com/flowers/setid.mat Begin to download .... Download finished
The loss value printed in the log is the current step, and the metric is the average value of previous step. Epoch 1/200
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/utils.py:77: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working return (isinstance(seq, collections.Sequence) and /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/nn/layer/norm.py:648: UserWarning: When training, we now always track global mean and variance. "When training, we now always track global mean and variance.")
step 40/49 [=======================>......] - loss: 1.4217 - acc: 0.5318 - ETA: 12:06 - 81s/st
二、paddle模型转化为onnx模型
使用百度的paddle2onnx工具进行转化
!pip insyall paddle2onnx
!pip install onnx
安装好之后，按照官方示例，加载模型，然后使用paddle.onnx.export接口进行转化，现在流行动态图，我们直接使用动态图操作方式，让静态图静静的躺在历史的尘埃里吧！！！
###注意###
模型输入需要设定batch为1，地平线开发板写死了入口，指定输入为4D,且batch只能为1，因此转化的时候需要设置如下形式：
input_spec = paddle.static.InputSpec(shape=[1, 3, 224, 224], dtype='float32', name='image')
In [ ]
import paddle import paddle2onnximport onnxfrom paddle.vision.models import MobileNetV2print(onnx.__version__,paddle.__version__,paddle2onnx.__version__)# 实例化模型mobilenetv2 = MobileNetV2(num_classes=102)#封装并加载模型model = paddle.Model(mobilenetv2) model.load(path='./save_model/0')#把模型从封装的Model中剥离出来net = model.network# 将模型设置为推理状态net.eval()# 定义输入数据input_spec = paddle.static.InputSpec(shape=[1, 3, 224, 224], dtype='float32', name='image')# ONNX模型导出# enable_onnx_checker设置为True，表示使用官方ONNX工具包来check模型的正确性，需要安装ONNX（pip install onnx）paddle.onnx.export(net, 'mobilenet_v2', input_spec=[input_spec], opset_version=10, enable_onnx_checker=True)
三、onnx模型校验
成功得到onnx模型之后，我们需要把onnx模型转化为板端部署文件 地平线提供了天工开物工具链，可以有效的对onnx模型进行验证并转化
部署第一步，使用工具链对onnx模型进行校验,校验不通过的模型目前无法上板部署，需要调整算子或提交地平线开发人员进行算子支持升级！！
工具链可在地平线生态社区获取，大家可以自行搜索！
开发板支持模型列表：
In [ ]
#!/usr/bin/env sh# Copyright (c) 2020 Horizon Robotics.All Rights Reserved.## The material in this file is confidential and contains trade secrets# of Horizon Robotics Inc. This is proprietary information owned by# Horizon Robotics Inc. No part of this work may be disclosed,# reproduced, copied, transmitted, or used in any way for any purpose,# without the express written permission of Horizon Robotics Inc.set -ex
cd $(dirname $0) || exit
model_type="onnx"onnx_model="../../../01_common/modelzoo/mapper/classification/mobilenet/mobilenet_v2.onnx"output="./mobilenet_checker.log"march="bernoulli2"hb_mapper checker --model-type ${model_type} \
--model ${onnx_model} \
--output ${output} --march ${march}四、onnx模型转化为板端推理文件
使用地平线的开发工具天公开物，开发套件内包含各种写好的脚本，可以快速实现校验与推理，节省开发者时间！
In [ ]
#!/bin/bash# Copyright (c) 2020 Horizon Robotics.All Rights Reserved.## The material in this file is confidential and contains trade secrets# of Horizon Robotics Inc. This is proprietary information owned by# Horizon Robotics Inc. No part of this work may be disclosed,# reproduced, copied, transmitted, or used in any way for any purpose,# without the express written permission of Horizon Robotics Inc.set -e -v
cd $(dirname $0) || exit
config_file="./mobilenet_config.yaml"model_type="caffe"# build modelhb_mapper makertbin --config ${config_file} \
--model-type ${model_type}In [ ]
# Copyright (c) 2020 Horizon Robotics.All Rights Reserved.## The material in this file is confidential and contains trade secrets# of Horizon Robotics Inc. This is proprietary information owned by# Horizon Robotics Inc. No part of this work may be disclosed,# reproduced, copied, transmitted, or used in any way for any purpose,# without the express written permission of Horizon Robotics Inc.# 模型转化相关的参数# model conversion related parametersmodel_parameters: # Caffe浮点网络数据模型文件 # the model file of floating-point Caffe neural network data onnx_model: '../../../01_common/modelzoo/mapper/classification/mobilenet/mobilenet_v2.onnx' # Caffe网络描述文件 # the file describes the structure of Caffe neural network #prototxt: '../../../01_common/modelzoo/mapper/classification/mobilenet/mobilenet_deploy.prototxt' # 适用BPU架构 # the applicable BPU architecture march: "bernoulli2" # 指定模型转换过程中是否输出各层的中间结果，如果为True，则输出所有层的中间输出结果， # --------------------------------------------------------------------------------------- # specifies whether or not to dump the intermediate results of all layers in conversion # if set to True, then the intermediate results of all layers shall be dumped layer_out_dump: False # 用于设置上板模型输出的layout, 支持NHWC和NCHW, 输入None则使用模型默认格式 # --------------------------------------------------------------------- # is used for specifying the layout of model output on dev board # both NHWC and NCHW layouts are supported, # if input is None then the default layout of the model will be used output_layout: None # 日志文件的输出控制参数， # debug输出模型转换的详细信息 # info只输出关键信息 # warn输出警告和错误级别以上的信息 # ---------------------------------------------------------------------------------------- # output control parameter of log file(s), # if set to 'debug', then details of model conversion will be dumped # if set to 'info', then only important imformation will be dumped # if set to 'warn', then information ranked higher than 'warn' and 'error' will be dumped log_level: 'debug' # 模型转换输出的结果的存放目录 # the directory in which model conversion results are stored working_dir: 'model_output' # 模型转换输出的用于上板执行的模型文件的名称前缀 # model conversion generated name prefix of those model files used for dev board execution output_model_file_prefix: 'mobilenetv2_pd'# 模型输入相关参数, 若输入多个节点, 则应使用';'进行分隔, 使用默认缺省设置则写None# -------------------------------------------------------------------------# model input related parameters,# please use ";" to seperate when inputting multiple nodes,# please use None for default settinginput_parameters: # (可不填) 模型输入的节点名称, 此名称应与模型文件中的名称一致, 否则会报错, 不填则会使用模型文件中的节点名称 # ------------------------------------------------------------------------------------------------- # (it's OK to leave blank) node name of model input, # it shall be the same as the name of model file, otherwise an error will be reported, # the node name of model file will be used when left blank #input_name: data # 网络实际执行时，输入给网络的数据格式，包括 nv12/rgbp/bgrp/yuv444_128/gray/featuremap, # 如果输入的数据为yuv444_128, 模型训练用的是bgrp，则hb_mapper将自动插入YUV到BGRP(NCHW)转化操作 # ------------------------------------------------------------------------------------------ # the data formats to be passed into neural network when actually performing neural network # whose values includes: nv12/rgbp/bgrp/yuv444_128/gray/featuremap # note that hb_mapper will automatically convert YUV into BGRP(NCHW) # if the input data is yuv444_128 while BGRP is used in model training input_type_rt: 'yuv444_128' # 网络训练时输入的数据格式，可选的值为rgbp/bgrp/gray/featuremap/yuv444_128 # --------------------------------------------------------------------- # the data formats in network training # available options include: rgbp/bgrp/gray/featuremap/yuv444_128 input_type_train: 'bgrp' # 模型网络的输入大小, 以'x'分隔, 不填则会使用模型文件中的网络输入大小，否则会覆盖模型文件中输入大小 # input_shape: '' # 网络输入的预处理方法，主要有以下几种： # no_preprocess 不做任何操作 # data_mean 减去通道均值mean_value # data_scale 对图像像素乘以data_scale系数 # data_mean_and_scale 减去通道均值后再乘以scale系数 # ------------------------------------------------------------------------------------------- # the input size of model network, seperated by 'x' # note that the network input size of model file will be used if left blank # otherwise it will overwrite the input size of model file # input_shape: '' # preprocessing methods of network input, consist of the follwing: # 'no_preprocess' indicates that no preprocess will be made # 'data_mean' indicates that to minus the channel mean, i.e. mean_value # 'data_scale' indicates that image pixels to multiply data_scale ratio # 'data_mean_and_scale' indicates that to multiply scale ratio after channel mean is minused norm_type: 'data_mean_and_scale' # 图像减去的均值, 如果是通道均值，value之间必须用空格分隔 # -------------------------------------------------------------------------- # the mean value minused by image # note that values must be seperated by space if channel mean value is used mean_value: 103.94 116.78 123.68 # 图像预处理缩放比例，如果是通道缩放比例，value之间必须用空格分隔 # --------------------------------------------------------------------------- # scale value of image preprocess # note that values must be seperated by space if channel scale value is used scale_value: 0.017calibration_parameters: # 模型量化的参考图像的存放目录，图片格式支持Jpeg、Bmp等格式，输入的图片 # 应该是使用的典型场景，一般是从测试集中选择20~100张图片，另外输入 # 的图片要覆盖典型场景，不要是偏僻场景，如过曝光、饱和、模糊、纯黑、纯白等图片 # 若有多个输入节点, 则应使用';'进行分隔 # ----------------------------------------------------------------------- # the directory where reference images of model quantization are stored # image formats include JPEG, BMP etc. # should be classic application scenarios, usually 20~100 images are picked out from test datasets # in addition, note that input images should cover typical scenarios # and try to avoid those overexposed, oversaturated, vague, # pure blank or pure white images # use ';' to seperate when there are multiple input nodes cal_data_dir: './calibration_data_bgrp' # 如果输入的图片文件尺寸和模型训练的尺寸不一致时，并且preprocess_on为true， # 则将采用默认预处理方法(skimage resize)， # 将输入图片缩放或者裁减到指定尺寸，否则，需要用户提前把图片处理为训练时的尺寸 # ---------------------------------------------------------------------------------- # In case the size of input image file is different from that of in model training # and that preprocess_on is set to True, # shall the default preprocess method(skimage resize) be used # i.e., to resize or crop input image into specified size # otherwise user must keep image size as that of in training in advance preprocess_on: False # 模型量化的算法类型，支持kl、max，通常采用KL即可满足要求 # ------------------------------------------------------------ # types of model quantization algorithms # kl and max are supported, usually kl works calibration_type: 'kl'# 编译器相关参数# compiler related parameterscompiler_parameters: # 编译策略，支持bandwidth和latency两种优化模式; # bandwidth以优化ddr的访问带宽为目标； # latency以优化推理时间为目标 # ------------------------------------------------------------------------------------------ # compilation strategy, there are 2 available optimization modes: 'bandwidth' and 'lantency' # the 'bandwidth' mode aims to optimize ddr access bandwidth # while the 'lantency' mode aims to optimize inference duration compile_mode: 'latency' # 设置debug为True将打开编译器的debug模式，能够输出性能仿真的相关信息，如帧率、DDR带宽占用等 # ----------------------------------------------------------------------------------- # the compiler's debug mode will be enabled by setting to True # this will dump performance simulation related information # such as: frame rate, DDR bandwidth usage etc. debug: False # 编译模型指定核数，不指定默认编译单核模型, 若编译双核模型，将下边注释打开即可 # ------------------------------------------------------------------------------------- # specifies number of cores to be used in model compilation # as default, single core is used as this value left blank # please delete the "# " below to enable dual-core mode when compiling dual-core model # core_num: 2 # ------------------------------------------------------------------------------------- # 优化等级可选范围为O0~O3 # O0不做任何优化, 编译速度最快，优化程度最低, # O1-O3随着优化等级提高，预期编译后的模型的执行速度会更快，但是所需编译时间也会变长。 # 推荐用O2做最快验证 # ---------------------------------------------------------------------------- # optimization level ranges between O0~O3 # O0 indicates that no optimization will be made # the faster the compilation, the lower optimization level will be # O1-O3: as optimization levels increase gradually, model execution, after compilation, shall become faster # while compilation will be prolonged # it is recommended to use O2 for fastest verification optimize_level: 'O3'
In [ ]
完成onnx模型->地平线开发板模型，转化得到的bin文件很小，只有2.4M，小了一个量级以上
五、部署与测试
继续使用天公开物工具链，调用上板测试模块
In [ ]
1.构建应用：# parallel_process_num的设置，参考您的cpu配置，如果不设置这个环境变量，默认为单进程cd samples/04_detection/01_yolov2/runtime_arm
sh 01_build.sh2.数据预处理# parallel_process_num的设置，参考您的cpu配置，如果不设置这个环境变量，默认为单进程export PARALLEL_PROCESS_NUM=${parallel_process_num}
sh 02_preprocess.sh
处理好的图像将会用于后续的评测eval3.将构建好的应用传到开发板
sh 03_scp_to_board.sh ${board_ip}
执行这个命令，会将构建好的应用，通过scp，传输到开发板的 /userdata/samples/mobilenet 目录下。
若要执行单张图片的infer, 则可通过下面的代码操作远程登录开发板进行执行
ssh root@${board_ip}
cd /userdata/samples/yolov2
sh dev_board_01_infer.sh4.执行评测
sh 04_eval.sh ${board_ip}
该脚本会将图片传输至板上进行运行和评测, 此过程将持续很久.5.执行性能测试
sh 05_perf.sh ${board_ip}
同时该操作也可从开发板上单独完成, 则可通过下面的代码操作远程登录开发板进行执行
ssh root@${board_ip}
cd /userdata/samples/yolov2
sh dev_board_03_perf.sh
————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————
最终结果：性能测试，10张图像仅用了34.01ms，速度经得起检验，精度损失极少
I0101 10:41:40.973683 1899 simple_example.cc:172] Whole process statistics:count:10, duration:34.01ms, min:3.342ms, max:3.823ms, average:3.35562ms, fps:294.031/s
, Infer stage statistics:count:10, duration:33.869ms, min:3.33ms, max:3.794ms, average:3.34312ms, fps:295.255/s
, Post process stage statistics:count:10, duration:0.125ms, min:0.01ms, max:0.027ms, average:0.011ms, fps:80000/s
