当前位置:   article > 正文

【CANN训练营第三季】AI开发入门_1. 使用tensorflow1.15实现lenet网络的minist手写数字识别。硬件平台不限,可

1. 使用tensorflow1.15实现lenet网络的minist手写数字识别。硬件平台不限,可

1、购买云服务器

在这里插入图片描述

在这里插入图片描述

2、快速体验

参考
https://gitee.com/ascend/samples/tree/master/cplusplus/level2_simple_inference/2_object_detection/YOLOV3_coco_detection_picture

# 为了方便下载,在这里直接给出原始模型下载及模型转换命令,可以直接拷贝执行。也可以参照上表在modelzoo中下载并手工转换,以了解更多细节。   
 
cd $HOME/samples/cplusplus/level2_simple_inference/2_object_detection/YOLOV3_coco_detection_picture/model     
wget https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/003_Atc_Models/AE/ATC%20Model/Yolov3/yolov3.caffemodel
wget https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/003_Atc_Models/AE/ATC%20Model/Yolov3/yolov3.prototxt
wget https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/models/YOLOV3_coco_detection_picture/aipp_nv12.cfg
atc --model=yolov3.prototxt --weight=yolov3.caffemodel --framework=0 --output=yolov3 --soc_version=Ascend310 --insert_op_conf=aipp_nv12.cfg

cd $HOME/samples/cplusplus/level2_simple_inference/2_object_detection/YOLOV3_coco_detection_picture/scripts    
bash sample_build.sh
bash sample_run.sh
  • 1
  • 2
  • 3
  • 4
  • 5
  • 6
  • 7
  • 8
  • 9
  • 10
  • 11

在这里插入图片描述
在这里插入图片描述
在这里插入图片描述
在这里插入图片描述

3、图像分类

参考:https://gitee.com/ascend/samples/tree/master/cplusplus/level2_simple_inference/1_classification/resnet50_imagenet_classification

4、结业考核

大题1、

1、

目标:

让学员动手编译和运行应用程序,并根据应用程序的参考实现及文档等理解其实现过程。

实战场景:

下载基于Caffe ResNet-50网络实现

图片分类(仅推理)样例应用的源码,并参考Readme成功编译、运行应用,体验基础推理过程。

评分细则:

总分30分:

使用atc工具转换模型,提供转换命令及转换成功的截图。(10分)

使用转换后的模型,重新编译运行样例应用(基于Caffe
ResNet-50网络实现图片分类(仅推理)),提交成功编译运行应用的截图。(10分)

总结实战过程中遇到的问题及解决方法,并提交总结。(5分)

优化样例应用,包括优化代码逻辑、优化代码注释、补充代码注释等,提交优化后的源码、优化思路说明、优化位置说明。(5分)

解答:

使用atc工具转换模型,提供转换命令及转换成功的截图。(10分)
cd ~/samples/cplusplus/level2_simple_inference/1_classification/resnet50_imagenet_classification
mkdir caffe_model
cd caffe_model
wget https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/003_Atc_Models/AE/ATC%20Model/resnet50/resnet50.prototxt
wget https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/003_Atc_Models/AE/ATC%20Model/resnet50/resnet50.caffemodel
cd ..
atc --model=caffe_model/resnet50.prototxt --weight=caffe_model/resnet50.caffemodel --framework=0 --output=model/resnet50 --soc_version=Ascend310 --input_format=NCHW --input_fp16_nodes=data --output_type=FP32 --out_nodes=prob:0
cd data
wget https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/models/aclsample/dog1_1024_683.jpg
wget https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/models/aclsample/dog2_1024_683.jpg
python3 ../script/transferPic.py
  • 1
  • 2
  • 3
  • 4
  • 5
  • 6
  • 7
  • 8
  • 9
  • 10
  • 11

模型转换成功的截图如下
在这里插入图片描述
在这里插入图片描述
执行python脚本需要安装对应的包

 which pip
     /home/HwHiAiUser/.local/bin/pip
 sudo ln -s /home/HwHiAiUser/.local/bin/pip /usr/local/python3.7.5/bin/pip 

这样pip安装就装在的python3.7.5


  • 1
  • 2
  • 3
  • 4
  • 5
  • 6
  • 7

在这里插入图片描述

使用转换后的模型,重新编译运行样例应用(基于Caffe ResNet-50网络实现图片分类(仅推理)),提交成功编译运行应用的截图。(10分)
export DDK_PATH=$HOME/Ascend/ascend-toolkit/latest
export NPU_HOST_LIB=$DDK_PATH/runtime/lib64/stub
cd ~/samples/cplusplus/level2_simple_inference/1_classification/resnet50_imagenet_classification
mkdir -p build/intermediates/host
cd build/intermediates/host
cmake ../../../src -DCMAKE_CXX_COMPILER=g++ -DCMAKE_SKIP_RPATH=TRUE
make
  • 1
  • 2
  • 3
  • 4
  • 5
  • 6
  • 7

遇到了一个小错误
fatal error: AclLiteApp.h: No such file or directory
#include “AclLiteApp.h”
^~~~~~~~~~~~~~
compilation terminated.
CMakeFiles/main.dir/build.make:110: recipe for target ‘CMakeFiles/main.dir/sample_process.cpp.o’ failed
make[2]: *** [CMakeFiles/main.dir/sample_process.cpp.o] Error 1
CMakeFiles/Makefile2:67: recipe for target ‘CMakeFiles/main.dir/all’ failed
make[1]: *** [CMakeFiles/main.dir/all] Error 2
Makefile:129: recipe for target ‘all’ failed
make: *** [all] Error 2

find / -name AclLiteApp.h 2> /dev/null
  • 1

原来是环境变量没配置好,需要重新配环境变量,也就是上面这段命令的前两行需要根据实际情况来。
在这里插入图片描述

cd $HOME/samples/cplusplus/level2_simple_inference/1_classification/resnet50_imagenet_classification/out
chmod +x main
./main

  • 1
  • 2
  • 3
  • 4

在这里插入图片描述
这样就运行成功了。

总结实战过程中遇到的问题及解决方法,并提交总结。(5分)

总结内容如上。

优化样例应用,包括优化代码逻辑、优化代码注释、补充代码注释等,提交优化后的源码、优化思路说明、优化位置说明。(5分)

优化为自动一键下载模型(模型初始化)
下载照片并进行分类即可(使用外链)
mkdir data0 && cd data0

同时把标号显示为查表后得到的单词结果。

这里我想列举文件夹下所有文件,需要用到
a.cpp:5:10: fatal error: io.h: No such file or directory
5 | #include <io.h>
| ^~~~~~
compilation terminated.
在ubuntu下,我写了如下代码获取文件夹中文件名称。

//头文件
#include <iostream>
#include <sys/types.h>
#include <dirent.h>
#include <vector>
#include <string.h>
 
using namespace std;
 
 
void GetFileNames(string path,vector<string>& filenames)
{
    DIR *pDir;
    struct dirent* ptr;
    if(!(pDir = opendir(path.c_str())))
        return;
    while((ptr = readdir(pDir))!=0) {
        if (strcmp(ptr->d_name, ".") != 0 && strcmp(ptr->d_name, "..") != 0)
            filenames.push_back(path + "/" + ptr->d_name);
    }
    closedir(pDir);
}
 
int main() {
    vector<string> file_name;
    string path = "/usr";
 
    GetFileNames(path, file_name);
 
    for(int i = 0; i <file_name.size(); i++)
    {
        cout<<file_name[i]<<endl;
    }
 
    return 0;
}
  • 1
  • 2
  • 3
  • 4
  • 5
  • 6
  • 7
  • 8
  • 9
  • 10
  • 11
  • 12
  • 13
  • 14
  • 15
  • 16
  • 17
  • 18
  • 19
  • 20
  • 21
  • 22
  • 23
  • 24
  • 25
  • 26
  • 27
  • 28
  • 29
  • 30
  • 31
  • 32
  • 33
  • 34
  • 35
  • 36

结合修改可以得到
我们要想办法把这个修改为动态获取data下文件
在这里插入图片描述
我写的脚本如下:

if [ ! -d "data" ];then
    mkdir data
else
    rm -rf data
    mkdir data
    # echo "data文件夹已经存在"
fi

cd data

read -p "input a pic:" pic
wget ${pic}
python3 ../script/transferPic.py
cd ..

if [ ! -d "caffe_model" ];then
    mkdir caffe_model
else
    echo "caffe_model文件夹已经存在"
fi

cd caffe_model
FILE1=resnet50.prototxt

if test -f "$FILE1"; then
    echo "$FILE1 exist"
else
    wget https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/003_Atc_Models/AE/ATC%20Model/resnet50/resnet50.prototxt
fi

FILE2=resnet50.caffemodel
if test -f "$FILE2"; then
    echo "$FILE2 exist"
else
    wget https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/003_Atc_Models/AE/ATC%20Model/resnet50/resnet50.caffemodel
fi

echo "模型初始化完成"

cd ..

export DDK_PATH=$HOME/Ascend/ascend-toolkit/latest
export NPU_HOST_LIB=$DDK_PATH/runtime/lib64/stub
cd ~/samples/cplusplus/level2_simple_inference/1_classification/resnet50_imagenet_classification
if [ ! -d "build" ];then
    mkdir -p build/intermediates/host
else
    rm -rf build
    mkdir -p build/intermediates/host
fi

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

cd $HOME/samples/cplusplus/level2_simple_inference/1_classification/resnet50_imagenet_classification/out
chmod +x main
./main

  • 1
  • 2
  • 3
  • 4
  • 5
  • 6
  • 7
  • 8
  • 9
  • 10
  • 11
  • 12
  • 13
  • 14
  • 15
  • 16
  • 17
  • 18
  • 19
  • 20
  • 21
  • 22
  • 23
  • 24
  • 25
  • 26
  • 27
  • 28
  • 29
  • 30
  • 31
  • 32
  • 33
  • 34
  • 35
  • 36
  • 37
  • 38
  • 39
  • 40
  • 41
  • 42
  • 43
  • 44
  • 45
  • 46
  • 47
  • 48
  • 49
  • 50
  • 51
  • 52
  • 53
  • 54
  • 55
  • 56
  • 57
  • 58
  • 59

最后却出现了一些小问题,很是恼火。
终于发现 是tranferPIC.py的问题,最后成功定制:
定制点:可以指定路径或者URL链接完成分类。
在这里插入图片描述
运行
bash test.sh #这是我写的一个脚本
比如我们可以在网页上复制我们想要分类的图像连接
在这里插入图片描述
在这里插入图片描述
粘贴到我的这个脚本里面,回车即可。
开源地址:
https://gitee.com/qmckw/AI-classification-ascend-AUTO-get-URL

大题2、

替换为resnet101

  • 目标

    基于前面所体验的样例,在样例源码的基础上,实现如下定制开发,完成相关模型适配并推理成功。

  • 实战场景

    基于Caffe ResNet-50网络实现图片分类(仅推理)样例,更换为ResNet-101分类模型。更换同类模型时,由于模型的输入、输出相似,所以源码基本可以复用,只需更换模型即可。

  • 定制点说明如下:

    1. 模型转换:

      下载ResNet-101模型(resnet101_tf.pb),放到resnet50_imagenet_classification样例的caffe_model目录(表示原始模型的存放路径),执行以下命令转换模型,在output参数处指定的model目录下获取resnet101_tf.om:

使用atc工具转换模型,提供转换命令及转换成功的截图。(5分)
cd samples/cplusplus/level2_simple_inference/1_classification/resnet50_imagenet_classification/
wget "https://ascend-repo-modelzoo.obs.cn-east-2.myhuaweicloud.com/model/ATC%20Resnet101(FP16)%20from%20TensorFlow%20-%20Ascend310/zh/1.1/ATC%20Resnet101(FP16)%20from%20TensorFlow%20-%20Ascend310.zip"

unzip 'ATC Resnet101(FP16) from TensorFlow - Ascend310.zip' -d resnet101-ascend310

mkdir caffe_model
cp resnet101-ascend310/Resnet101_for_TensorFlow/resnet101_tf.pb ./caffe_model/resnet101_tf.pb

 atc --model=caffe_model/resnet101_tf.pb --framework=3 --output=model/resnet101_tf --output_type=FP32 --soc_version=Ascend310 --input_shape="input:1,224,224,3" --log=info
  • 1
  • 2
  • 3
  • 4
  • 5
  • 6
  • 7
  • 8
  • 9

在这里插入图片描述

模型的基本介绍如下:

  • 输入数据

    输入数据大小数据类型数据排布格式
    input224 x 224RGB_FP32NHWC
  • 输出数据

    输出数据大小(batch_size x 类别数)数据类型数据排布格式
    resnet_v1_101/predictions/Softmaxbatch_size x 1000FLOAT32ND

数据格式一般只能是“NHWC”或“NCHW”,默认是前者。“NHWC“表示[batch, height, width, channels],如果是”NCHW“表示[batch, channels, height, width]。

  1. 生成测试数据

    进入resnet50_imagenet_classification样例的script目录,修改transferPic.py脚本中的如下内容,将float16改为float32:

    img = img.astype(“float16”)

    修改后:

    img = img.astype(“float32”)
    在这里插入图片描述
    切换到“resnet50_imagenet_classification样例目录/data“目录下,执行transferPic.py脚本,将*.jpg转换为*.bin,同时将图片从1024683的分辨率缩放为224224(需要修改py文件)。在“resnet50_imagenet_classification样例目录/data“目录下生成2个*.bin测试文件。

    python3 …/script/transferPic.py

在这里插入图片描述

  1. 调用AscendCL接口(例如aclmdlLoadFromFileWithMem接口)加载ResNet-101模型:

    在src/sample_process.cpp文件中定制代码。
    在这里插入图片描述
    只改个文件名。
    在这里插入图片描述

参考Caffe ResNet-50网络实现图片分类(仅推理)样例的readme,重新编译并运行。
rm -rf build
mkdir -p build/intermediates/host
cd build/intermediates/host
export DDK_PATH=$HOME/Ascend/ascend-toolkit/latest
export NPU_HOST_LIB=$DDK_PATH/runtime/lib64/stub
cmake ../../../src -DCMAKE_CXX_COMPILER=g++ -DCMAKE_SKIP_RPATH=TRUE
make
  • 1
  • 2
  • 3
  • 4
  • 5
  • 6
  • 7
使用转换后的模型、定制后的代码,重新编译运行应用,提交成功编译运行应用的截图。(10分)

在这里插入图片描述
运行成功,我们可以看到resnet101用的value是大于1的,而resnet50貌似都是小于1的value.

总结实战过程中遇到的问题及解决方法,并提交总结。(5分)

总结如上。
我完成后开源代码:
https://gitee.com/qmckw/resnet101_imagenet_classification

大题3、

2023年1月3日我看到题目有所修订,原先是用tf实现的:

3、使用Pytorch实现LeNet网络的minist手写数字识别。
硬件平台不限,可以基于windows或者linux系统,尽量给出整个过程的截图,并在最后给出loss或者accuracy运行结果,提供打印loss和accuracy日志,给出截图。【20分】
参考链接Github:https://github.com/allegrofb/LeNet.git
解答:

git clone https://github.com/allegrofb/LeNet.git 
cd LeNet

  • 1
  • 2
  • 3

在这里插入图片描述
我们可以看到,该脚本训练两轮。
在这里插入图片描述
在这里插入图片描述

原题目:

  1. 使用Tensorflow1.15实现LeNet网络的minist手写数字识别。硬件平台不限,可以基于window或者linux系统,尽量给出整个过程的截图,并在最后给出Loss或者accuracy运行结果。 参考链接Gitee或Github。【15分】
    提交样式:步骤xx截图xx
  2. 采用课程中学习到的手工或自动迁移方式,将上述脚本迁移到昇腾AI处理器上,不要求执行训练,仅需提交迁移后的脚本。【15分】
  3. 将上述迁移好的LeNet网络使用minist数据集在ModelArts平台上正常跑通。

评分明细:

a. Pycharm控制台界面有正常训练日志打印,给出截图。【5分】

b. 将最终训练的模型权重文件(训练步数不限)保存在OBS上,给出截图。【2分】

c. 给出当前训练的CANN运行日志截图,给出截图。【3分】

解答:
参考https://gitee.com/lai-pengfei/LeNet

git clone https://gitee.com/lai-pengfei/LeNet
conda activate tensorflow1 # 我创建的一个tf环境
pip install tensorflow==1.15

  • 1
  • 2
  • 3
  • 4

中间遇到了一个报错,找不到tensorflow的tutorial模块
修改一下import的方式

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
  • 1
  • 2

这样就在本地顺利地跑起来了

  1. 使用Tensorflow1.15实现LeNet网络的minist手写数字识别。硬件平台不限,可以基于window或者linux系统,尽量给出整个过程的截图,并在最后给出Loss或者accuracy运行结果。 参考链接Gitee或Github。【15分】
    在这里插入图片描述
    跑通之后,迁移脚本
  2. 采用课程中学习到的手工或自动迁移方式,将上述脚本迁移到昇腾AI处理器上,不要求执行训练,仅需提交迁移后的脚本。【15分】
cd ..
git clone https://gitee.com/ascend/tensorflow.git
cd tensorflow/convert_tf2npu #进入工具包
pip3 install pandas
pip3 install xlrd==1.2.0
pip3 install openpyxl
pip3 install tkintertable
pip3 install google_pasta
python3 main.py -i ../../LeNet

  • 1
  • 2
  • 3
  • 4
  • 5
  • 6
  • 7
  • 8
  • 9
  • 10

在这里插入图片描述
在这里插入图片描述
在这里插入图片描述
查看报告可知,迁移已经成功。
在这里插入图片描述
并且成功得到了输出脚本
下面使用pycharm在modelarts上训练,配置如图
3. 将上述迁移好的LeNet网络使用minist数据集在ModelArts平台上正常跑通。

评分明细:

a. Pycharm控制台界面有正常训练日志打印,给出截图。【5分】

在这里插入图片描述

在这里插入图片描述
在这里插入图片描述

这个时候出现了一个小错误
在这里插入图片描述
data作相应修改

mnist = input_data.read_data_sets("/home/ma-user/modelarts/user-job-dir/code/MNIST_data/", one_hot = True)
  • 1

在这里插入图片描述
迭代设置为5000次,很快就得到了运行结果
在这里插入图片描述
在OBS得到了对应的checkpoint文件
b. 将最终训练的模型权重文件(训练步数不限)保存在OBS上,给出截图。【2分】

在这里插入图片描述

c. 给出当前训练的CANN运行日志截图,给出截图。【3分】
在这里插入图片描述
在这里插入图片描述

在这里插入图片描述
在这里插入图片描述
在这里插入图片描述

2、开发者自己选择非昇腾社区ModelZoo的Tensorflow模型,迁移成功并在ModelArts平台上训练跑通。

评分细则:

  1. 迁移成功,提交迁移后的脚本。【5分】

  2. 在ModelArts平台训练跑通,给出相关日志截图。【5分】

友情提醒:

  1. 昇腾社区ModelZoo的Tensorflow模型列表可单击LINK,并通过如下筛选条件查找:

在这里插入图片描述

  1. 没有迁移或训练成功也没关系,可在ModelZoo仓提issue,也能酌情给分。

我们选择Retina-VesselNet迁移

git clone https://github.com/DeepTrial/Retina-VesselNet.git
git clone https://gitee.com/ascend/tensorflow.git
cd tensorflow/convert_tf2npu #进入工具包
pip3 install pandas
pip3 install xlrd==1.2.0
pip3 install openpyxl
pip3 install tkintertable
pip3 install google_pasta
python3 main.py -i ../../Retina-VesselNet-v2
  • 1
  • 2
  • 3
  • 4
  • 5
  • 6
  • 7
  • 8
  • 9

我们利用迁移工具,可以发现已经成功迁移
在这里插入图片描述

然后因为我们这个是tf2.0,那么我们可以选用启智平台安装了tf2.0的ascend镜像
训练情况如图,可以运行。
在这里插入图片描述

大题4、

自动迁移

声明:本文内容由网友自发贡献,不代表【wpsshop博客】立场,版权归原作者所有,本站不承担相应法律责任。如您发现有侵权的内容,请联系我们。转载请注明出处:https://www.wpsshop.cn/w/寸_铁/article/detail/791655
推荐阅读
相关标签
  

闽ICP备14008679号