Homework 1-4 learning notes

No.1

Use a voucher to buy ECS

1. Accept shared image

2. Use coupons
Click on the coupon to use it immediately —— Buy Elastic servers —— Buy on demand

Image select shared image


bandwidth 100
Set account password

Quick experience stay ECS Run up

establish secssion With root user With SSH Log in to ECS Switch to HwHIUser Compile under the user Run the application

Use the public network here IP And then root user User name is root The other is myself Setting the name is useless

Switch HwHiUser user

Switch to this user In this folder

Running example

Enter the code cloud warehouse

according to readme To operate

Enter the model Directory Then turn the model

No.2

Linux Basic operation

The principle of deep learning

AI Full stack architecture

No.3

Image classification application development

Model transformation

Model transformation tools ATC

Original model ：TensorFlow/Caffe
Single operator ：json
Model transformation mode ： Command line 、MindStudio
The resulting ：om file

op name: Operator name
There is a list of relevant supported operators in the document

Use of model transformation tools ：

choice prototxt And caffemodel Model

310： Reasoning chip 910： Training pictures
load configuration: Import the configuration file that was successfully converted last time
Eye picture ： Model visualization tool Searchable correlation operators

FP16: Semi precision floating point
select output node: Right click node choice Mark, You can output the specified operator
The size here is the image input required by the model

end ,finish


above json Files are related configuration files , You can import

To configure AIPP

AI pre-processing
Data preprocessing ：AIPP（ Model transformation phase configuration , Solidify into the model ） DVPP
BT.601： Standard video definition format
BT.609: Standard high definition format
The size here is AIPP Input the desired picture
AIPP Can achieve the function of image enhancement The alignment edge of the image is eliminated （ No basis , Don't understand, ）, Color gamut conversion of pictures 、 Image cropping 、 Picture normalization ： mean value 、 minimum value 、 variance

Single operator model file configuration

1. Model transformation ：


2. Write profile

input output Parametric model :


Example ：

ACL executive summary

ascend computing language Rising computing language
effect ： Computational acceleration



Implicitly create Implicitly release According to create Explicit release

Main interface calling process and practice

You should have such a process architecture in mind before you start

ACL Of helloworld Code

Configure the build file ：

• Use cmake Method compilation
  There are changes ：（ Other copy and paste ）
  
  // The source file you want to compile
  
  // The path where the header file is placed

Introduction to the environment ：

AI1 virtual machine
compile ：



make compile

Get into out And create an empty file

And then execute
see ACL Run log ：
Environmental Science ：


here host journal plog-10036( Process number )-2021…（ Time stamp ）

here device journal

The level of logging Log printing unit Time stamp Source file Print the content
A. The level of logging B. Name of the module that generated the log C. Process name D. The file that triggers the log generation and the corresponding line number E. Specific log contents of each module

ACL Run resource management

It's mainly about Device、Context、Stream Conduct management
It is suggested to create three resources
The relationship between the three ：

Device management

Pay attention to the release sequence ：stream-context-device

Example ：


Compile and run ：
：

Context management

establish ：

The destruction ：




Example ：

Stream management

establish ：

The destruction ：

Example ：

No.5

Attention to big homework ：
Replace the picture with the same name