Toc
  1. quickly see
  2. ModelArts use
    1. guide
    2. local code
      1. online training
      2. PyCharm
      3. VS Code
    3. local dataset
    4. local development with SDK
      1. python package
  3. CANN
  4. Ascend in Huawei Cloud
  5. MindSpore installation
  6. Euler OS
  7. Mind Studio
  8. MindX SDK
Toc
0 results found
MINGG
首页 归档 分类 标签 books 链接 关于
ascend
2021/09/10

Created: 2021-06-07 00:29:11

Modified: 2021-09-10 10:31

use ascend

quickly see

python -c "import mindspore;print(mindspore.__version__)"
import mindspore
from mindspore import context
from mindspore.context import ParallelMode
# data parallel - training data divided
# after BP: AllReduce
context.set_auto_parallel_context(parallel_mode=ParallelMode.DATA_PARALLEL)
# op level model parallel
self.matmul=ops.MatMul().shard(((1,2),(2,1)))
c=self.matmul(A,B)
context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL)

context.set_auto_parallel_context(parallel_mode=ParallelMode.AUTO_PARALLEL)

self.matmul=ops.MatMul.set_strategy({[4,1],[1,1]})
c=self.matmul(A,B)
context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL)
# pipline level model parallel
# I think it is better that all of the pieces of pipline occupy the same time of calculation.
self.subnet1 = subnet().pipeline_stages(0)
self.subnet2 = subnet().pipeline_stages(1)
x = self.subnet1(x)
x = self.subnet2(x)
context.set_auto_parallel_context(pipeline_stages=2)
# Rematerialization -- reduce the ocupation of training data in DRAM
# there is gradient value produced in DRAM too
# there are weight and other parameter(like EMA for optimizer) stored in SRAM
x = self.net1(x)
self.net1.recompute() # ?
x = self.net2(x)
x = self.net1(x)
self.net1.recompute()
x = self.net2(x)
# data parallel - parameter shard -- divide parameter of inner layers into SRAMs
# befor FP: AllGether
# after BP: ReduceScatter
# example: fusion every 10 layers of transformer for OpenAI - gpt3
for i in range(config.num_layers):
    self.blocks.append(Block(config, i+1)).set_comm_fusion(int(i/10)+2)
context.set_auto_parallel_context(enable_parallel_optimizer=True,parallel_mode=ParallelMode.DATA_PARALLEL)

ModelArts use

guide

local code in PC -> cloud code in obs -> my ModelArts algorithm /algorithm subscribed in AI Gallery -> Training Jobs ->

local dataset in PC/cloud dataset in AI Gallery -> cloud dataset in obs-> Training Jobs ->

-> output in Training Jobs -> output in obs

local code

according to PyCharm with the Plugin of ModelArts:

online training

local code in PC -> cloud code in obs -> my ModelArts algorithm -> Training Jobs -> output in Training Jobs -> output in obs

=> local code in PC -> the whole job & log in PyCharm

image-20210728201155839

image-20210728203122068

what you need to install are 3 items, refer to https://support.huaweicloud.com/bestpractice-modelarts/modelarts_10_0021.html:

PyCharm

create and open one on Notebook -> remote configure on PyCharm -> restart and wait -> enter remote interpreter environment opening a script(connecting to remote host) -> upload files(synch and auto-upload files) -> wait a while(after index files) and then open requirements.txt and install dependency ->

open terminal -> cd work/pangu_alpha -> ls -> python src/preprocess.py -> bash myrun.sh -> debug train.py（trial）

image-20210806145821878

image-20210806145852764

Update the whitelist of Notebook -> connect in PyCharm-> (tool-deployment) -> (change remote host) enter remote interpreter environment opening a script -> upload files(synch and auto-upload files) -> wait a while(after index files) and then open requirements.txt and install dependency ->

open terminal -> cd work/PanguGit-> ls -> python src/preprocess.py -> bash myrun.sh -> (configure first, like python interpreter)run train.py

Test, refer to https://bbs.huaweicloud.com/forum/thread-133615-1-1.html:

ssh ma-user@dev-modelarts-cnnorth4.huaweicloud.com -p 31261 -i .ssh/KeyPair-remote-dev.pem
python -c "import mindspore;print(mindspore.__version__)"
exit

image-20210909202354600

VS Code

refer to 配置本地IDE（VSCode手动连接）.

Firstly, make sure you use VSCode 1.57.1 with Extensions installed including Python and Remote-SSH, create a Notebook with SSH and prepare a pair of SSH key of Huawei Cloud.

Host {name of Notebook1}
     HostName 114.116.242.242
     Port {Port of Notebook1}
     User ma-user
     IdentityFile ~\.ssh\KeyPair-xxx.pem
     UserKnownHostsFile=/dev/null
     StrictHostKeyChecking no
Host {name of Notebook2}
     HostName dev-modelarts-cnnorth4.huaweicloud.com
     Port {Port of Notebook2}
     User ma-user
     IdentityFile ~\.ssh\KeyPair-xxx.pem
     UserKnownHostsFile=/dev/null
     StrictHostKeyChecking no

image-20210909222413407

Ensure consistency between Nootbook and config file(Host, Port, path of key pair in particular) -> make sure Remote-SSH configured -> Update the whitelist of IP and open Notebook -> click the button of Connect to Host in New Window in vscode and wait a while(download VS Code server and copy it to host with scp) -> use VS Code server like a New one (install Extensions of VS code -> manage Folder in Cloud(drag and drop files from PC) -> Select Interpreter of Python (need to open a py file first in vscode) -> open New Terminal -> run and debug) -> close the Romote Connection and stop Notebook:

image-20210909235814519

image-20210909232000263

Note that you may want to disable the software update:

image-20210909222324964

local dataset

local dataset in PC -> cloud dataset in obs->

use OBS Browser+ or obsutil to make it easily.



local development with SDK

use ModelArts SDK, refer to https://support.huaweicloud.com/sdkreference-modelarts/modelarts_04_0002.html

there is only python package version.

python package

from modelarts.session import Session
session = Session(access_key='***',secret_key='***', project_id='***', region_name='cn-north-4')
session.obs.upload_file(src_local_file='/home/ma-user/file1.txt', dst_obs_dir='obs://bucket-name/dir1/')
session.obs.upload_dir(src_local_dir='/home/ma-user/', dst_obs_dir='obs://bucket-name/dir1/')
session.obs.download_file(src_obs_file="obs://bucket-name/dir1/file1.txt", dst_local_dir="/home/ma-user/")
session.obs.download_dir(src_obs_dir="obs://bucket-name/dir1/", dst_local_dir="/home/ma-user/")

Note: it seems not to support \ of path of Windows. and there is something wrong with dir operation?

CANN

refer to CANN社区版安装指南(5.0.3.alpha002),

  1. Prepare your devices including both hardware and software ahead. For example, you can [rent a server and run an Image in Huawei Cloud](#Ascend in Huawei Cloud).

  2. (optional)Create and substitute with a User for installation or running, or use an image with user created in the step 1.

  3. Install and check dependency of apt and pip and firmware with root right, or use an image with dependency and firmware installed in the step 1. What’s more,

    image-20210917045150015

    Besides, add environmental variables:

    # in ~/.bashrc
    export LD_LIBRARY_PATH=/usr/local/python3.7.5/lib:$LD_LIBRARY_PATH
    export PATH=/usr/local/python3.7.5/bin:$PATH

    For firmware of Driver,

    # dowaload before
    chmod +x A*
    su
    ./A300-3010-npu-driver_21.0.2_linux-x86_64.run --full
    # install in /usr/local/Ascend
    reboot

  1. Prepare the .run package of CANN to the device and then install it. In insider version you can use command line to download and install it. The default installation path is /usr/local/Ascend or ~/Ascend.

  2. validated the environmental variables of CANN

    source ${install_path}/bin/setenv.bash
    source ~/.bashrc
    # farther
    # 存在多个python版本时，需要指定为python3.7.5的安装路径，请根据实际情况修改
    export PATH=/usr/local/python3.7.5/bin:$PATH
    # for non-root user
    export install_path=~/Ascend
    # for root user
    export install_path=/usr/local/Ascend
    # ascend driver
    export a_driver_path=/usr/local/Ascend/driver
    # 加载Driver的so文件存在路径
    export LD_LIBRARY_PATH=/usr/local/Ascend/driver/lib64/common/:/usr/local/Ascend/driver/lib64/driver:$LD_LIBRARY_PATH
    # Runtime使用到的so文件存在路径
    # 如环境中存在Driver，则配置如下。
    export LD_LIBRARY_PATH=${install_path}/runtime/lib64:$LD_LIBRARY_PATH
    # 如环境中不存在Driver，则配置如下。
    export LD_LIBRARY_PATH=${install_path}/runtime/lib64:$LD_LIBRARY_PATH:${install_path}/runtime/lib64/stub
    # Compiler使用的python路径
    export PYTHONPATH=${install_path}/compiler/python/site-packages:$PYTHONPATH
    # 如果用户环境存在多个python3版本，则指定使用python3.7.5版本，python3.7.5安装路径请根据实际情况进行替换；同时设置ATC工具可执行文件所在路径
    export PATH=/usr/local/python3.7.5/bin:${install_path}/compiler/ccec_compiler/bin:${install_path}/compiler/bin:$PATH
    # Compiler的库文件存在路径
    export LD_LIBRARY_PATH=${install_path}/compiler/lib64:${install_path}/compiler/lib64/plugin/opskernel:${install_path}/compiler/lib64/plugin/nnengine:${install_path}/compiler/lib64/stub
    # FwkPlugin使用的python路径
    export PYTHONPATH=${install_path}/fwkplugin/python/site-packages:$PYTHONPATH
    # Toolkit使用的python路径
    export PYTHONPATH=${install_path}/toolkit/python/site-packages:$PYTHONPATH
    export TOOLCHAIN_HOME=${install_path}/toolkit
    # 算子库根目录
    export ASCEND_OPP_PATH=${install_path}/opp
    #AI CPU组件的安装路径
    export ASCEND_AICPU_PATH=${install_path}/

Ascend in Huawei Cloud

refer to https://support.huaweicloud.com/ecs/index.html

image-20210917010717644

image-20210917010810298

image-20210917011100973

image-20210917011014249

image-20210917011256290

image-20210917011900989

image-20210917012005473

remote login with a software in PC, like Termius in Mac OS, MobaXterm in Windows,

articles about MobaXterm: MobaXterm技巧,

image-20210917015223036

Note that it is supported to drag and drop files from PC.

MindSpore installation

Make sure dependency installed first, refer to MindSpore安装指南, ML environment.

Taking the installation of MindSpore 1.2 on Ubuntu-x86(GPU CUDA 10.1) for example, only one line refer to MindSpore安装指南:

pip install https://ms-release.obs.cn-north-4.myhuaweicloud.com/1.2.1/MindSpore/gpu/ubuntu_x86/cuda-10.1/mindspore_gpu-1.2.1-cp37-cp37m-linux_x86_64.whl --trusted-host ms-release.obs.cn-north-4.myhuaweicloud.com -i https://pypi.tuna.tsinghua.edu.cn/simple
# Successfully installed asttokens-2.0.5 astunparse-1.6.3 cffi-1.14.5 decorator-5.0.9 easydict-1.9 mindspore-gpu-1.2.1 mpmath-1.2.1 numpy-1.21.0 packaging-20.9 pillow-8.2.0 protobuf-3.17.3 psutil-5.8.0 pycparser-2.20 pyparsing-2.4.7 scipy-1.7.0 six-1.16.0 sympy-1.8

Check your installation:

pip show mindspore-gpu
# Name: mindspore-gpu
# Version: 1.2.1
# Summary: MindSpore is a new open source deep learning training/inference framework that could be used for mobile, edge and cloud scenarios.
# Home-page: https://www.mindspore.cn
# Author: The MindSpore Authors
# Author-email: contact@mindspore.cn
# License: Apache 2.0
# Location:
# Requires: cffi, sympy, protobuf, scipy, setuptools, astunparse, easydict, wheel, psutil, packaging, numpy, decorator, asttokens, pillow
# Required-by: 
python -c "import mindspore;print(mindspore.__version__)"
# 1.2.1

Test 1:

import numpy as np
import mindspore.context as context
import mindspore.nn as nn
from mindspore import Tensor
from mindspore.ops import operations as P

context.set_context(mode=context.GRAPH_MODE, device_target="GPU")

class Mul(nn.Cell):
    def __init__(self):
        super(Mul, self).__init__()
        self.mul = P.Mul()

    def construct(self, x, y):
        return self.mul(x, y)

x = Tensor(np.array([1.0, 2.0, 3.0]).astype(np.float32))
y = Tensor(np.array([4.0, 5.0, 6.0]).astype(np.float32))

mul = Mul()
print(mul(x, y))
# [4.0, 10.0, 18.0]

Test 2:

import numpy as np
from mindspore import Tensor
import mindspore.ops as ops
import mindspore.context as context

context.set_context(device_target="GPU")
x = Tensor(np.ones([1,3,3,4]).astype(np.float32))
y = Tensor(np.ones([1,3,3,4]).astype(np.float32))
print(ops.add(x, y))
# [[[ 2.  2.  2.  2.],
#     [ 2.  2.  2.  2.],
#     [ 2.  2.  2.  2.]],

#     [[ 2.  2.  2.  2.],
#     [ 2.  2.  2.  2.],
#     [ 2.  2.  2.  2.]],

#     [[ 2.  2.  2.  2.],
#     [ 2.  2.  2.  2.],
#     [ 2.  2.  2.  2.]]]

Test 3:

refer to 体验快速入门,

git clone https://gitee.com/mindspore/docs.git
cd docs
git checkout r1.2

Euler OS

# check your system
uname -a
# Linux notebook-c4f2050c-c720-44cd-a6bc-ca40ad2e7bfd 3.10.0-862.14.1.5.h328.eulerosv2r7.x86_64 #1 SMP Mon Jul 22 00:00:00 UTC 2019 x86_64 x86_64 x86_64 GNU/Linux

# or
uname -r
# 3.10.0-862.14.1.5.h328.eulerosv2r7.x86_64

# unavailable
export HELLO="Hello!"

Mind Studio

refer to 昇腾CANN训练营-应用营第一讲——开发&运行环境部署 (0:46:36).

MindX SDK

https://education.huaweicloud.com/courses/course-v1:HuaweiX+CBUCNXA042+Self-paced/about?isAuth=0&cfrom=hwc

image-20210916225911222

image-20210916230133104

image-20210916230226497

image-20210916230406927

7个API

img

打赏
支付宝
微信
shared documents
ubuntu
本文作者：MINGG
永久链接：http://mingg2333.top/2021/09/10/hiascend/
版权声明：本文首发于MINGG的博客，转载请注明出处！