【CANN训练营】体验基于Caffe ResNet-50网络实现图片分类实践操作

CANN版本：CANN-8.0.RC3.alpha001。

JeffDingAI

788人浏览 · 2026-01-07 11:07:38

JeffDingAI · 2026-01-07 11:07:38 发布

实例功能

很简单的一个实例，功能就是一个实现图片分类的功能，然后拓展实现以下

将一张YUV420SP格式的图片编码为*.jpg格式的图片。
将两张*.jpg格式的解码成两张YUV420SP NV12格式的图片，缩放，再进行模型推理，分别得到两张图片的推理结果后，处理推理结果，输出最大置信度的类别标识以及top5置信度的总和。
将两张*.jpg格式的解码成两张YUV420SP NV12格式的图片，抠图，再进行模型推理，分别得到两张图片的推理结果后，处理推理结果，输出最大置信度的类别标识以及top5置信度的总和。
将两张*.jpg格式的解码成两张YUV420SP NV12格式的图片，抠图贴图，再进行模型推理，分别得到两张图片的推理结果后，处理推理结果，输出最大置信度的类别标识以及top5置信度的总和。
将YUV420SP NV12格式的图片（分辨率8192*8192）缩放，得到4000*4000。

环境及环境版本介绍

NPU：Ascend910（32GB）

CANN版本：CANN-8.0.RC3.alpha001

开始实践

创建conda环境

conda create -n cann_demo python=3.8 -y
conda activate cann_demo

安装CANN

wget https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/Milan-ASL/Milan-ASL%20V100R001C20SPC703/Ascend-cann-toolkit_8.0.0.alpha003_linux-aarch64.run
bash Ascend-cann-toolkit_8.0.0.alpha003_linux-aarch64.run --full

激活环境变量

source /home/ma-user/Ascend/ascend-toolkit/set_env.sh

下载体验代码仓

git clone -b v0.3-8.0.0.alpha003 https://gitee.com/Ascend/samples

进入示例文件夹

cd samples/cplusplus/level2_simple_inference/1_classification/vpc_jpeg_resnet50_imagenet_classification

获取ResNet-50原始模型

下载模型CAFFE文件

cd caffe_model
wget https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/003_Atc_Models/AE/ATC%20Model/resnet50/resnet50.caffemodel
wget https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/003_Atc_Models/AE/ATC%20Model/resnet50/resnet50.prototxt
cd ..

安装依赖

pip install decorator attrs psutil sympy scipy

转换模型

atc --model=caffe_model/resnet50.prototxt --weight=caffe_model/resnet50.caffemodel --framework=0 --soc_version=Ascend910 --insert_op_conf=caffe_model/aipp.cfg --output=model/resnet50_aipp

准备测试图片

cd data
wget https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/models/aclsample/dvpp_vpc_8192x8192_nv12.yuv
wget https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/models/aclsample/persian_cat_1024_1536_283.jpg
wget https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/models/aclsample/wood_rabbit_1024_1061_330.jpg
wget https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/models/aclsample/wood_rabbit_1024_1068_nv12.yuv
cd ..

编译运行

安装依赖

conda install -c conda-forge cmake
conda install -c conda-forge binutils

创建目录

mkdir -p build/intermediates/host

设置环境变量

source /home/ma-user/Ascend/ascend-toolkit/set_env.sh
export DDK_PATH=$HOME/Ascend/ascend-toolkit/latest
export NPU_HOST_LIB=$DDK_PATH/runtime/lib64/stub

生成编译文件

cd build/intermediates/host
cmake ../../../src -DCMAKE_CXX_COMPILER=g++ -DCMAKE_SKIP_RPATH=TRUE
make

运行

设置main文件权限为可运行

cd ../../../out
chmod +x main

将两张*.jpg格式的解码成两张YUV420SP NV12格式的图片，缩放，再进行模型推理，分别得到两张图片的推理结果

./main 0

运行结果

[INFO]  ./main param, param represents a vpc feature and must be set
[INFO]  start check result fold:./result
[INFO]  make directory successfully.
[INFO]  check result success, fold exist
[INFO]  acl init success
[INFO]  set device 0 success
[INFO]  create context success
[INFO]  create stream success
[INFO]  get run mode success
[INFO]  dvpp init resource success
[INFO]  load model ../model/resnet50_aipp.om success
[INFO]  create model description success
[INFO]  create model output success
[INFO]  model input width 224, input height 224
[INFO]  -------------------------------------------
[INFO]  start to process picture:../data/persian_cat_1024_1536_283.jpg
[INFO]  call JpegD
[INFO]  call vpcResize
[INFO]  Process dvpp success
[INFO]  create model input success
[INFO]  model execute success
[INFO]  destroy model input success
[INFO]  result : classType[283], top1[0.969727], top5[0.979855]
[INFO]  -------------------------------------------
[INFO]  start to process picture:../data/wood_rabbit_1024_1061_330.jpg
[INFO]  call JpegD
[INFO]  call vpcResize
[INFO]  Process dvpp success
[INFO]  create model input success
[INFO]  model execute success
[INFO]  destroy model input success
[INFO]  result : classType[331], top1[0.895508], top5[1.000134]
[INFO]  -------------------------------------------
[INFO]  unload model success, modelId is 1
[INFO]  destroy model description success
[INFO]  destroy model output success
[INFO]  execute sample success
[INFO]  end to destroy stream
[INFO]  end to destroy context
[INFO]  end to reset device 0
[INFO]  end to finalize acl

将两张*.jpg格式的解码成两张YUV420SP NV12格式的图片，抠图，再进行模型推理，分别得到两张图片的推理结果。

./main 1

运行结果

[INFO]  ./main param, param represents a vpc feature and must be set
[INFO]  start check result fold:./result
[INFO]  check result success, fold exist
[INFO]  acl init success
[INFO]  set device 0 success
[INFO]  create context success
[INFO]  create stream success
[INFO]  get run mode success
[INFO]  dvpp init resource success
[INFO]  load model ../model/resnet50_aipp.om success
[INFO]  create model description success
[INFO]  create model output success
[INFO]  model input width 224, input height 224
[INFO]  -------------------------------------------
[INFO]  start to process picture:../data/persian_cat_1024_1536_283.jpg
[INFO]  call JpegD
[INFO]  call vpcCrop
[INFO]  Process dvpp success
[INFO]  create model input success
[INFO]  model execute success
[INFO]  destroy model input success
[INFO]  result : classType[283], top1[0.996094], top5[0.999629]
[INFO]  -------------------------------------------
[INFO]  start to process picture:../data/wood_rabbit_1024_1061_330.jpg
[INFO]  call JpegD
[INFO]  call vpcCrop
[INFO]  Process dvpp success
[INFO]  create model input success
[INFO]  model execute success
[INFO]  destroy model input success
[INFO]  result : classType[330], top1[0.859863], top5[1.000106]
[INFO]  -------------------------------------------
[INFO]  unload model success, modelId is 1
[INFO]  destroy model description success
[INFO]  destroy model output success
[INFO]  execute sample success
[INFO]  end to destroy stream
[INFO]  end to destroy context
[INFO]  end to reset device 0
[INFO]  end to finalize acl

将两张*.jpg格式的解码成两张YUV420SP NV12格式的图片，抠图贴图，再进行模型推理，分别得到两张图片的推理结果。

./main 2

运行结果

[INFO]  ./main param, param represents a vpc feature and must be set
[INFO]  start check result fold:./result
[INFO]  check result success, fold exist
[INFO]  acl init success
[INFO]  set device 0 success
[INFO]  create context success
[INFO]  create stream success
[INFO]  get run mode success
[INFO]  dvpp init resource success
[INFO]  load model ../model/resnet50_aipp.om success
[INFO]  create model description success
[INFO]  create model output success
[INFO]  model input width 224, input height 224
[INFO]  -------------------------------------------
[INFO]  start to process picture:../data/persian_cat_1024_1536_283.jpg
[INFO]  call JpegD
[INFO]  call vpcCropAndPaste
[INFO]  Process dvpp success
[INFO]  create model input success
[INFO]  model execute success
[INFO]  destroy model input success
[INFO]  result : classType[283], top1[0.431885], top5[0.751892]
[INFO]  -------------------------------------------
[INFO]  start to process picture:../data/wood_rabbit_1024_1061_330.jpg
[INFO]  call JpegD
[INFO]  call vpcCropAndPaste
[INFO]  Process dvpp success
[INFO]  create model input success
[INFO]  model execute success
[INFO]  destroy model input success
[INFO]  result : classType[330], top1[0.685059], top5[0.969410]
[INFO]  -------------------------------------------
[INFO]  unload model success, modelId is 1
[INFO]  destroy model description success
[INFO]  destroy model output success
[INFO]  execute sample success
[INFO]  end to destroy stream
[INFO]  end to destroy context
[INFO]  end to reset device 0
[INFO]  end to finalize acl

将一张YUV420SP格式的图片编码为*.jpg格式的图片。

./main 3

运行结果

[INFO]  ./main param, param represents a vpc feature and must be set
[INFO]  start check result fold:./result
[INFO]  check result success, fold exist
[INFO]  acl init success
[INFO]  set device 0 success
[INFO]  create context success
[INFO]  create stream success
[INFO]  get run mode success
[INFO]  dvpp init resource success
[INFO]  start to jpege picture ../data/wood_rabbit_1024_1068_nv12.yuv
[INFO]  end to destroy stream
[INFO]  end to destroy context
[INFO]  end to reset device 0
[INFO]  end to finalize acl

将一张分辨率为8192*8192的YUV420SP格式的图片缩放至4000*4000。

./main 4

运行结果

[INFO]  ./main param, param represents a vpc feature and must be set
[INFO]  start check result fold:./result
[INFO]  check result success, fold exist
[INFO]  acl init success
[INFO]  set device 0 success
[INFO]  create context success
[INFO]  create stream success
[INFO]  get run mode success
[INFO]  dvpp process 8k resize begin
[INFO]  dvpp init resource success
[INFO]  dvpp process 8k resize success
[INFO]  end to destroy stream
[INFO]  end to destroy context
[INFO]  end to reset device 0
[INFO]  end to finalize acl

整体运行结果

执行可执行文件成功后，同时会在main文件同级的result目录下生成结果文件，便于后期查看。结果文件如下：

dvpp_output_0：persian_cat_1024_1536_283.jpg：图片经过缩放或抠图或抠图贴图之后的结果图片。
dvpp_output_1：wood_rabbit_1024_1061_330.jpg：图片经过缩放或抠图或抠图贴图之后的结果图片。
model_output_0：persian_cat_1024_1536_283.jpg：图片的模型推理结果，二进制文件。
model_output_0.txt：persian_cat_1024_1536_283.jpg：图片的模型推理结果，txt文件。
model_output_1：wood_rabbit_1024_1061_330.jpg：图片的模型推理结果，二进制文件。
model_output_1.txt：wood_rabbit_1024_1061_330.jpg：图片的模型推理结果，txt文件。
jpege_output_0.jpg：wood_rabbit_1024_1068_nv12.yuv：图片结果编码后的结果图片。
dvpp_vpc_4000x4000_nv12.yuv：dvpp_vpc_8192x8192_nv12.yuv：图片缩放后的结果图片。

昇腾开源生态专区

昇腾计算产业是基于昇腾系列（HUAWEI Ascend）处理器和基础软件构建的全栈 AI计算基础设施、行业应用及服务，https://devpress.csdn.net/organization/setting/general/146749包括昇腾系列处理器、系列硬件、CANN、AI计算框架、应用使能、开发工具链、管理运维工具、行业应用及服务等全产业链

更多推荐

HarmonyOS 6 自定义人脸识别模型10：基于MindSpore Lite框架的自定义人脸识别功能实现

昇腾开源生态专区

DeepSeek 崩了 13 小时，不是故障，是 V4 在换引擎

昇腾开源生态专区

体系结构论文（107）：AscendOptimizer: Episodic Agent for Ascend NPU Operator Optimization

摘要：本文提出AscendOptimizer系统，针对华为Ascend NPU的AscendC算子优化难题，通过双阶段交替优化方法实现性能提升。系统将算子拆分为host侧tiling program和device侧kernel program：Stage I采用进化搜索优化tiling策略，利用硬件反馈探索可行解空间；Stage II通过"优化回退"机制从优质kernel反向构