ArcGIS pro2.8 deep learning environment configuration based on rtx30 graphics card

Problem description

For reasons I don't know ,ArcGIS Pro2.8 Does not support RTX30 The graphics card , In the deep learning model training, the model accuracy is very low ,loss Equivalent to nan The situation of , Specific to see ESRI Discussion post ：https://github.com/Esri/deep-learning-frameworks/issues/17
The questions summarized in the post are as follows ：

Solutions

Get to the point , To configure argis pro2.8 Deep learning environment to support RTX30 The idea of graphics card is ：
1、 install CUDA11 And the corresponding cudnn;
2、arcgis pro2.8 Deep learning environment library package replacement .

1、 install CUDA11+cudnn

First step ： Check the type of computer graphics card

The second step ： download cuda11 Install the package and install

The blogger's graphics card is 11.4, But installed 11.1 Of cuda, This is based on arcgis pro Deep learning library package version consideration , Finally, verify cuda Installation successful .
CUDA11.1 Download address ：https://developer.nvidia.com/cuda-11.1.1-download-archive

The third step ： Install the corresponding version of cudnn And configure environment variables

cudnn Download address ：https://developer.nvidia.com/zh-cn/cudnn
Just refer to the blog configuration above ：https://blog.csdn.net/A496608119/article/details/112855153

To configure ArcGIS Pro2.8 Deep learning environment

First step ： Configure deep learning environment arcgispro-py3

Installation reference ESRI Github：https://github.com/Esri/deep-learning-frameworks/blob/master/README.md?rmedium=links_esri_com_b_d&rsource=https%3A%2F%2Flinks.esri.com%2Fdeep-learning-framework-install
download ArcGIS Pro2.8 Deep learning environment configurator ：

Download and run MSI file , After completion of operation ,ArcGIS Pro2.8 You can bring it with you arcgispro-py3 Implement deep learning tools in the environment .

When the operation is completed, open ArcGIS Pro2.8, Execute training model tool , Successful execution means successful environment configuration .

The second step ： clone arcgispro-py3 Environment to deepleanring

open ArcGIS Pro Of python Interactive terminal
Enter the command ：conda create -n deeplearning --clone arcgispro-py3
After cloning your own environment , Enter the command ：activate deepleanring
After activating the new environment , The library package will be replaced in this environment .

The third step ： Replace deeplearning Library packages in the environment

First post the list of Library packages I have configured ：
(deeplearning) C:\WINDOWS\system32>conda list

The replaced library package should be downloaded ESRI Library package provided , Address ：https://anaconda.org/esri/repo?page=1
Main replacement pytorch Associated library package , All replaced library packages can be downloaded directly tar.bz2 file , Library packages are not upgraded or degraded , Core library package pytorch Still using 1.4.0 edition ,ESRI Provides the basis for cuda11.2+cudnn8.1 Of pytorch：

ESRI The provided download source does not contain library packages , You can download it from Tsinghua source ：https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/win-64/
https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/win-64/
The downloaded library package adopts conda Command to install ：conda install ********.tar.bz2（*** Is the local path of the library package file ）.
Last input ：proswap deeplearning, take ArcGIS Pro2.8 Of python Replace the environment with a new environment .

Step four ： modify ArcGIS Pro Small BUG

When training the model GBK error , Need modification C:\Program Files\ArcGIS\Pro\bin\Python\envs\deeplearning\Lib\site-packages\arcgis\learn_data.py The code in , In script with open() as f Add encoding=“utf-8”, After modification .