One 、 Related information

MindSpore_white_paperV1.1.pdf

Two 、 Important note

1. MindSpore Of Ascend Version and ascend-toolkit as well as driver There is a binding relationship between versions . Used on the shengteng platform MindSpore Depend on CANN, namely ascend-toolkit And driver Wait for the kit .

3、 ... and 、 Relevant concepts

3.1 ME

Mind Expression,MindSpore front end , It mainly completes the compilation tasks from user source code to calculation diagram 、 Control execution and context maintenance in training （ In non sinking mode configuration ）、 Dynamic graph （PyNative Pattern ） etc. .

3.2 MindData

MindSpore Data frame , Provide data loading 、 enhance 、 Data set management and visualization .

3.3 MindInsight

MindSpore Visualization components , Visual scalar 、 Images 、 Calculation diagram and model hyperparametric information .

3.4 AIR

Ascend Intermediate Representation, similar ONNX, It is an open file format designed for machine learning defined by Huawei , Better fit Ascend AI processor , Commonly used in Ascend 310 Perform reasoning tasks on .

AIR Format only MindSpore Ascend 910 Exportable , only Ascend 310 Reasonable , Before reasoning, use Ascend CANN in atc Tool for model transformation , Reasoning does not depend on MindSpore, Only Ascend CANN software package .

Ascend 310 It is energy-efficient and highly integrated for edge scenes AI processor , Support for MindIR Format and AIR Format model for reasoning .

3.5 MindIR

MindIR Full name MindSpore IR, A kind of Functional expression based on graph representation IR, Defines extensible The graph structure as well as operator IR Express , Store MindSpore Infrastructure . it Eliminates model differences between different backend , Generally used to perform reasoning tasks across hardware platforms .MindSpore Through unification IR The logical structure of the network and the properties of the operator are defined , take MindIR The format of the model file is decoupled from the hardware platform , Realization One training, many deployments .

MindIR The format can be MindSpore CPU、GPU、Ascend 910 export , Can run in GPU、Ascend 910、Ascend 310 On , There is no need to manually perform model transformation before reasoning , Reasoning requires installation MindSpore, call MindSpore C++ API Reasoning .

characteristic

1. MindIR As MindSpore The first mock exam , At the same time, the network structure and weight parameter values are stored . At the same time, it supports deployment to the cloud Serving And the end side Lite The platform performs reasoning tasks .

2. The same MindIR The file supports the deployment of various hardware forms ：

   • Cloud Serving Deployment reasoning ：MindSpore Training generation MindIR After the model file , It can be sent directly to MindSpore Serving load , Perform reasoning tasks , Without additional model transformation , Achieve Ascend、GPU、CPU And so on .
   • End side Lite Reasoning deployment ：MindIR It can be directly used for Lite Deployment use . At the same time, due to the end side lightweight demand , It provides model miniaturization and conversion functions , Support the original MindIR The model file is created by Protocol Buffers Format into FlatBuffers Format store , And lightweight network structure , To better meet the end side performance 、 Memory and other requirements .

Use scenarios

First use network definitions and Checkpoint File export MindIR Model file , Then perform reasoning tasks according to different needs .

• stay Ascend 310 Perform reasoning tasks on
• be based on MindSpore Serving Deploy inference Services
• End to side reasoning .

3.6 MindRecord

MindSpore A data format defined , Is an execution read 、 write in 、 Search and transform MindSpore Module of format data set .

3.7 MindSpore Lite

A lightweight deep neural network reasoning engine , Provided will be MindSpore The trained model has the function of reasoning on the end side .

3.8 PyNative Mode

MindSpore Dynamic patterns of patterns , Each operator in the neural network is issued and executed one by one , It is convenient for users to write and debug neural network models .

3.9 Graph Mode

MindSpore Static graph mode , Compile the neural network model into a whole picture , Then issue to execute , High performance .

3.10 MindArmour

MindSpore Security module , Through differential privacy 、 Antagonistic attack and defense and other technical means , Improve the confidentiality of the model 、 Integrity and availability , Prevent attackers from maliciously modifying the model or cracking the internals of the model , Steal the parameters of the model .

Four 、 Related introduction

1. MindSpore Back end support includes Ascend、GPU、CPU.

5、 ... and 、 Key tutorials

5.1 Installation plan （Windows）

MindStudio It can be installed separately Windows On . In the installation MindStudio Before you need stay Linux Install and deploy on the server Ascend-cann-toolkit Development kit , After the Windows Installation on MindStudio, Pass after installation Configure remote connection The way to build MindStudio Where Windows The server and Ascend-cann-toolkit Where the development kit is located Linux Server connection , Realize the whole process development function .

take MindStudio Installed in the Windows On the server ,Windows The server is a local environment ,Linux The server is a remote environment .

5.2 Installation plan （Linux）

MindStudio and Ascend-cann-toolkit have access to Linux The terminal of the native desktop on the server gnome-terminal Installation , It can also be in Windows Pass on server SSH Log in to Linux Install the server . because MindStudio Is a GUI Program , So in Windows Pass on server SSH Log in to Linux When installing the server , Integration is needed X server Of SSH terminal （ such as MobaXterm, The tool version needs to be v20.2 And above ）.

5.3 install MindStudio

Detailed installation tutorial , Please refer to the official website Installation guide .

5.4 install MindSpore

Get installation command

#  Create a conda Virtual environment and activate 
conda create -n mindsp python=3.7.5 -y
conda activate mindsp

#  install MindSpore
pip install https://ms-release.obs.cn-north-4.myhuaweicloud.com/1.7.0/MindSpore/cpu/x86_64/mindspore-1.7.0-cp37-cp37m-win_amd64.whl --trusted-host ms-release.obs.cn-north-4.myhuaweicloud.com -i https://pypi.tuna.tsinghua.edu.cn/simple

#  Verify that the installation was successful 
python -c "import mindspore;mindspore.run_check()"

Verify success

(mindsp-py37) D:\360Downloads\vision>python -c "import mindspore;mindspore.run_check()"
MindSpore version:  1.7.0
The result of multiplication calculation is correct, MindSpore has been installed successfully!

[[email protected] mindspore-lenet-demo]$ python3 -c "import mindspore;mindspore.run_check()"
[WARNING] ME(1564:281472980111376,MainProcess):2022-07-04-16:35:16.111.95 [mindspore/run_check/_check_version.py:267] Using custom Ascend 910 AI software package path, package version checking is skipped, please make sure Ascend 910 AI software package version is supported, you can reference to the installation guidelines https://www.mindspore.cn/install
[WARNING] ME(1564:281472980111376,MainProcess):2022-07-04-16:35:16.179.332 [mindspore/run_check/_check_version.py:368] Can not find driver so(need by mindspore-ascend), please check if you have set env LD_LIBRARY_PATH, you can reference to the installation guidelines https://www.mindspore.cn/install
MindSpore version: 1.5.0
The result of multiplication calculation is correct, MindSpore has been installed successfully!

5.5 MindStudio To configure SSH

SSH Connection management

5.6 install MindSpore Vision

install MindSpore Vision

Source code installation

#  Create a conda Virtual environment and activate 
conda create -n mindvision python=3.7.5 -y
conda activate mindvision

# git clone MindSpore Vision Warehouse 
git clone https://gitee.com/mindspore/vision.git
cd vision

#  Source code installation 
python setup.py install

#  Verify that the installation was successful 
python ./examples/classification/lenet/lenet_mnist_infer.py --data_url ./tests/st/classification/dataset/mnist/mnist.jpg --pretrained True --device_target CPU

Verify success

[WARNING] ME(25912:24800,MainProcess):2022-06-22-09:11:35.362.873 [mindspore\dataset\core\config.py:646] For Windows and MacOS we forbid shared mem function temporarily.
494592B [00:00, 1997626.51B/s]
{
    4: 'four'}

pip install

pip install mindvision

6、 ... and 、 Quick start ： Handwritten digit recognition

Follow the tutorial on the official website ： Quick start ： Handwritten digit recognition , Please refer to FAQ.

Mindspore：r1.7

MindSpore Vision：0.1.0

7、 ... and 、 model training

Follow the tutorial on the official website ： model training , Please refer to FAQ.

8、 ... and 、 stay ModelArts notebook Medium training

[[email protected] mindspore-lenet-demo]$ python3 lenet_demo.py
[WARNING] ME(842:281473601032208,MainProcess):2022-07-04-16:27:03.476.263 [mindspore/run_check/_check_version.py:267] Using custom Ascend 910 AI software package path, package version checking is skipped, please make sure Ascend 910 AI software package version is supported, you can reference to the installation guidelines https://www.mindspore.cn/install
[WARNING] ME(842:281473601032208,MainProcess):2022-07-04-16:27:03.980.652 [mindspore/run_check/_check_version.py:368] Can not find driver so(need by mindspore-ascend), please check if you have set env LD_LIBRARY_PATH, you can reference to the installation guidelines https://www.mindspore.cn/install
[WARNING] ME(842:281473601032208,MainProcess):2022-07-04-16:27:11.532.279 [mindspore/dataset/core/config.py:464] The shared memory is on, multiprocessing performance will be improved. Note: the required shared memory can't exceeds 80% of the available shared memory. You can reduce max_rowsize or reduce num_parallel_workers to reduce shared memory usage. [WARNING] ME(842:281473601032208,MainProcess):2022-07-04-16:27:16.404.060 [mindspore/dataset/core/config.py:464] The shared memory is on, multiprocessing performance will be improved. Note: the required shared memory can't exceeds 80% of the available shared memory. You can reduce max_rowsize or reduce num_parallel_workers to reduce shared memory usage.
[WARNING] SESSION(842,fffe827fc1e0,python3):2022-07-04-16:27:28.599.092 [mindspore/ccsrc/backend/session/ascend_session.cc:1806] SelectKernel] There are 9 node/nodes used reduce precision to selected the kernel!
Epoch:[  0/ 10], step:[ 1875/ 1875], loss:[0.138/0.138], time:27793.110 ms, lr:0.01000
Epoch time: 32266.532 ms, per step time: 17.209 ms, avg loss: 0.138
Epoch:[  1/ 10], step:[ 1875/ 1875], loss:[0.001/0.001], time:3550.877 ms, lr:0.01000
Epoch time: 3552.153 ms, per step time: 1.894 ms, avg loss: 0.001
Epoch:[  2/ 10], step:[ 1875/ 1875], loss:[0.017/0.017], time:3824.274 ms, lr:0.01000
Epoch time: 3825.681 ms, per step time: 2.040 ms, avg loss: 0.017
Epoch:[  3/ 10], step:[ 1875/ 1875], loss:[0.003/0.003], time:3692.203 ms, lr:0.01000
Epoch time: 3693.945 ms, per step time: 1.970 ms, avg loss: 0.003
Epoch:[  4/ 10], step:[ 1875/ 1875], loss:[0.025/0.025], time:3885.926 ms, lr:0.01000
Epoch time: 3887.030 ms, per step time: 2.073 ms, avg loss: 0.025
Epoch:[  5/ 10], step:[ 1875/ 1875], loss:[0.000/0.000], time:3726.956 ms, lr:0.01000
Epoch time: 3727.775 ms, per step time: 1.988 ms, avg loss: 0.000
Epoch:[  6/ 10], step:[ 1875/ 1875], loss:[0.042/0.042], time:3601.407 ms, lr:0.01000
Epoch time: 3602.245 ms, per step time: 1.921 ms, avg loss: 0.042
Epoch:[  7/ 10], step:[ 1875/ 1875], loss:[0.001/0.001], time:4478.034 ms, lr:0.01000
Epoch time: 4478.895 ms, per step time: 2.389 ms, avg loss: 0.001
Epoch:[  8/ 10], step:[ 1875/ 1875], loss:[0.006/0.006], time:4465.447 ms, lr:0.01000
Epoch time: 4466.262 ms, per step time: 2.382 ms, avg loss: 0.006
Epoch:[  9/ 10], step:[ 1875/ 1875], loss:[0.004/0.004], time:4313.395 ms, lr:0.01000
Epoch time: 4314.218 ms, per step time: 2.301 ms, avg loss: 0.004
[WARNING] SESSION(842,fffe827fc1e0,python3):2022-07-04-16:28:25.698.502 [mindspore/ccsrc/backend/session/ascend_session.cc:1806] SelectKernel] There are 4 node/nodes used reduce precision to selected the kernel!
{
    'accuracy': 0.9895833333333334}
[WARNING] ME(842:281473601032208,MainProcess):2022-07-04-16:28:33.318.388 [mindspore/dataset/core/config.py:464] The shared memory is on, multiprocessing performance will be improved. Note: the required shared memory can't exceeds 80% of the available shared memory. You can reduce max_rowsize or reduce num_parallel_workers to reduce shared memory usage.
[WARNING] SESSION(842,fffe827fc1e0,python3):2022-07-04-16:29:17.098.612 [mindspore/ccsrc/backend/session/ascend_session.cc:1806] SelectKernel] There are 4 node/nodes used reduce precision to selected the kernel!
Predicted: "[1 1 8 2 1 0]", Actual: "[1 1 8 2 1 0]"

Nine 、FAQ

Q：ImportError: cannot import name ‘Mnist’ from ‘mindvision.dataset’

Traceback (most recent call last):
  File "D:/MyDocuments/PyProjects/pydemo/lenet_demo.py", line 8, in <module>
    from mindvision.dataset import Mnist
ImportError: cannot import name 'Mnist' from 'mindvision.dataset' (D:\360Downloads\anaconda3\envs\mindsp-py37\lib\site-packages\mindvision-0.1.0-py3.7.egg\mindvision\dataset\__init__.py)

 The reason for the error ：
 Guide Package error 

from mindvision.dataset import Mnist
 Change it to 
from mindvision.classification.dataset import Mnist