Experimental reproduction of image classification (reasoning only) based on caffe resnet-50 network

Abstract ： This experiment is mainly based on Caffe ResNet-50 Network to achieve image classification （ Reasoning only ） For example , Learn how to do it when you already have a pre training model , Deploy the model to shengteng AI Reasoning on the processor .

This article is shared from Huawei cloud community 《【CANN Training camp 】【2022 Season two 】【 Novice class 】 be based on Caffe ResNet-50 Network to achieve image classification （ Reasoning only ） Experimental recurrence of 》, author ： StarTrek .

This experiment is mainly based on Caffe ResNet-50 Network to achieve image classification （ Reasoning only ） For example , Learn how to do it when you already have a pre training model , Deploy the model to shengteng AI Reasoning on the processor . The main tasks of this experiment are ：

1、 take Caffe ResNet-50 The model file of the network is converted to the adaptive shengteng AI Offline model of the processor （ * .om file ）;
2、 Load the om file , Yes 2 Zhang * .jpg Synchronous reasoning of pictures , After getting the reasoning results respectively , Then deal with the reasoning results , Output top5 Category identification of confidence ;
3、 Replace the test image with your own image and recompile and run .

One 、 The use of ECS

First of all, in order to better let us understand and learn relevant knowledge of shengteng , Huawei CANN The training camp has given each student a certain amount of cloud server resources , A brief summary of the use of ECS is as follows ：

The shared image provided this time is ：Ubuntu 18.04（ System ） + 5.1.RC2.alpha005（CANN Environment version ）

Landing in huaweiyun -> Access control -> apply ESC Elastic cloud server -> Get the elastic public network IP-> Application SSH Remote access to ECs -> According to the normal ubuntu The terminal command of the system is used to operate the server

Resource link ：

Remote terminal software recommendation MobaXterm：https://mobaxterm.mobatek.net/
How to use the software ：MobaXterm Detailed use of the tutorial _xuanying_china The blog of -CSDN Blog _mobaxterm

After entering the terminal, it is root user , Need to go to HwHiAiUser user

Instructions ：

	cd /home/
	su - HwHiAiUser

have access to ls perhaps ll Command to browse the existing files in the system at will .

Two 、 be based on Caffe ResNet-50 Network to achieve image classification （ Reasoning only ） experiment

After the preparatory work , Now you can start the formal experiment ！ You can follow the readme The steps described in the document are operated step by step , Here is just a simple description of the experimental steps and experimental diagram .

Warehouse resource link ：samples: CANN Samples - Gitee.com

1、 Download the sample code

clone sample The code in the warehouse

	git clone https://gitee.com/ascend/samples.git

Enter into resnet50_imagenet_classification In the folder of the samples

	cd samples/
	cd cplusplus/level2_simple_inference/1_classification/resnet50_imagenet_classification

2、 obtain ResNet-50 Pre training model

First create a folder to store the model , And open the folder

	mkdir caffe_model
	cd caffe_model

Download the pre training model directly through the command from the web page ： Weight file （resnet50.caffemodel） And model files （resnet50.prototxt）

	wegt https://modelzoo-train-atc.obs.cn-north-4.myhuaweicloud.com/003_Atc_Models/AE/ATC%20Model/resnet50/resnet50.prototxt
	wegt https://modelzoo-train-atc.obs.cn-north-4.myhuaweicloud.com/003_Atc_Models/AE/ATC%20Model/resnet50/resnet50.caffemodel

3、 Model transformation

The pre training model downloaded above needs to be used first ATC The model conversion tool converts the model into shengteng AI Offline model supported by processor （.om）
Official documents ATC Tool learning resources ： Shengteng community - Official website - shengteng Wanli Let intelligence be everywhere

technological process ：

First, switch to the sample directory （ Then the above operation is the upper level directory ）

	cd ../

Then run the model transformation tool

	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

open model The folder can be seen resnet50.om The conversion has been completed

4、 Download the test image

Get into data Folder , And download two ImageNet There are pictures of categories in the dataset

	cd ../data/
	wget https://c7xcode.obs.cn-north-4.myhuaweicloud.com/models/aclsample/dog1_1024_683.jpg
	wget https://c7xcode.obs.cn-north-4.myhuaweicloud.com/models/aclsample/dog2_1024_683.jpg

 

Two pictures of the test

5、 Picture format conversion

The downloaded picture is JPEG Format , It needs to be converted into a binary file suitable for the format required by the model input （.bin）
stay data Under the table of contents , perform transferPic.py Script , take *.jpg Convert to *.bin, At the same time, change the picture from 1024 * 683 The resolution of is scaled to 224 * 224.

	python3 ../script/transferPic.py

6、 Compile operation

Get into “cplusplus/level2_simple_inference/1_classification/resnet50_imagenet_classification” Sample catalog

	cd ../

Set temporary environment variables , Configures the header file and library file path that the compiler depends on

	export DDK_PATH=$HOME/Ascend/ascend-toolkit/latest
	export NPU_HOST_LIB=$DDK_PATH/acllib/lib64/stub

establish build Folder , Prepare to compile the code file

	mkdir -p build/intermediates/host

 

Get into build Folder , compile .cpp file

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

  It can be found in the sample directory （cplusplus/level2_simple_inference/1_classification/resnet50_imagenet_classification） Under the out Found the compiled executable in the folder main file , Next, let's run main file

	cd ../../../out/
	./main

3、 ... and 、 Replace image data for testing

Here I found two pictures casually on the Internet （ A goldfish label-1, A golden retriever label-207）, You can find several pictures at will , As long as it is JPG Format pictures can be , The best type of things in the picture is ImageNet Types in dataset labels .

1、 Add pictures and convert formats

You can download it directly to your local computer , Then drag the file directly to MobaXterm In the ECS file list of the software , For convenience, first find the previous one in the list data Folder , Then directly drag it to the folder .

Follow the previous operation （5、 Picture format conversion ） Re convert the image file format , stay data Run in directory

	python3 ../script/transferPic.py

2、 modify sample_process.cpp Source code , Add the path of the two new pictures

stay testFile Add the two newly added and converted ones to the list bin Format file path , The path can be written according to the first two , But don't forget the comma separation between different file paths

3、 Recompile source

Just follow the previous steps （6、 Compile operation ）, It should be noted that , If the server has been restarted , The previously set compilation temporary environment will be lost , You need to set the temporary environment variables again （ Configures the header file and library file path that the compiler depends on ）, Then you can go to build/intermediates/host Run directly under the file make Recompile, , The compiler will automatically compile the modified source file and overwrite .

4、 Reasoning

Rerun the compiled main File can （6、 The last step in compiling and running ）, Note the path of the file , Need to be in out Run under folder , because main The file is saved here .

Finally, we can see

The label is 1 It's goldfish , In the first picture, the probability of goldfish is 99.7070%
The label is 207 It's a golden retriever , In the second picture, the probability of golden retriever is 99.5605%

Click to follow , The first time to learn about Huawei's new cloud technology ~