Swin transformer object detection project installation tutorial

0 Basic environmental requirements

• pytorch
• CUDA
• Visual Studio
• anaconda

Be careful ：

• I 'm on my own pytorch The version is 1.9.1,CUDA The version is 11.1,anaconda The version is ：Anaconda3-2021.04-Windows-x86_64.exe,Visual Studio The version is Visual Studio Community 2019
• Another thing to note , Use python3.9 The following errors may appear in the version ：
  No matching distribution found for pycocotools_windows, in other words python3.9 Unable to install pycocotools_windows, So it's better to python Version changed to 3.6,3.7 or 3.8

1 install mmcv

Method 1 ： Offline installation ( I use this method )

• First step

clone mmcv Project to local （ Such as d:）

git clone -b v1.3.1 https://github.com/open-mmlab/mmcv.git

Swin Transformer Required by the project mvcc Version must be >=1.2.4 And <=1.4.0, Otherwise, an error will be reported . So I use mmcv The version is 1.3.1

• The second step ：

cd mmcv
pip install -r requirements.txt

• The third step ：

find cl.exe The path address of , Then add it to the environment variable , Usually cl.exe Your location is right here Visual Studio In the installation directory of

My cl.exe Position in ：

C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.28.29910\bin\Hostx64\x86

• Step four
  All the following commands need to be in Power Shell Command window execution ,Power Shell The position of the is shown in the red box in the figure below
  

open Power Shell, Enter the command as follows

set Path=c
cl

If the following results appear , Indicates that the environment variable is added successfully

• Step five

View the native GPU Calculate the force ：

You can use the following statement to view the GPU Calculate the force ：

import torch
print(' Of the current graphics card CUDA Calculate the force :',torch.cuda.get_device_capability())

You can also open the following link to find GPU Computing power
https://developer.nvidia.cn/zh-cn/cuda-gpus

My computer graphics card model is GTX1660TI, Cannot find your own in the table GPU model , So I chose something close to it RTX2060, The calculated force value is 7.5

• Step six

stay Power Shell Command window ：

$env:TORCH_CUDA_ARCH_LIST="7.5" # Estimate your own computer GPU Calculate the force to set 
$env:MMCV_WITH_OPS = 1
$env:MAX_JOBS = 4  #  According to what you can use CPU Set the number of cores 

• Step seven

# Switch to mmcv In the directory 
cd mmcv
# compile 
python setup.py build_ext 
# install 
python setup.py build_ext 

Method 2 ： Online installation
Be careful ： I just started with online installation , Although the installation was successful , But it's running later Swin Transformer Project demo when , The following error message appears ：

No module named ‘mmcv_custom‘

Only later , It was because I installed mmcv The version is 1.4.0, This version is too high , You need to reduce the version , But according to my cuda Version and torch Version can only be used 1.4.0 And above , So give up method two , Method 1 is adopted , And successfully installed 1.3.1 Version of ,

• First step

First look at your own torch Version and CUDA edition , Then open the mmcv Of github link ：https://github.com/open-mmlab/mmcv/blob/master/README.md, Check the configuration table below , Such as my torch The version is 1.9,CUDA The version is 11.1, So I choose the installation method shown in the red box below

• The second step ：

Open the arrow in the red box in the above figure , The details are as follows ：

• The third step ：

Then you need to open the link in the red box above ：https://download.openmmlab.com/mmcv/dist/cu111/torch1.9.0/index.html, Check whether there is a required version in this link , Because my system is windows System ,python The version is 3.8, And I have to choose 1.4.0 Up to , So the optional results at this time are shown in the following red box

Be careful ：

• Swin Transformer Required by the project mvcc Version must be ：>=1.2.4 And <=1.4.0, Otherwise, an error will be reported , For example, I installed it for the first time mvcc The version is 1.4.1, The following error message will appear
  

• Step four ;

Enter the following installation command on the command line ：

pip install mmcv-full== 1.4.0 -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.9.0/index.html

Be careful ： Here I download according to my version requirements , You can also download according to your specific computer situation , The method is similar to the above process

2 install mmdetection

mmdetection The version of must be the same as mmcv Versions match , Refer to the table below for details , Because of my mmcv The version is 1.3.1, therefore mmdetection The version of is set to 2.11.0

Enter the following four commands on the command line ：

git clone -b v2.11.0 https://github.com/open-mmlab/mmdetection.git
cd mmdetection
pip install -r requirements/build.txt
pip install -v -e . # or "python setup.py develop"

3 install apex

It is used for half precision training , It can save memory and speed up training

clone apex Project to local （ Such as d:）

git clone https://github.com/NVIDIA/apex

Execute the following command ：

cd apex
python setup.py install

4 install Swin-Transformer-Object-Detection

• clone mmcv Project to local （ Such as d:）

    git clone https://github.com/SwinTransformer/Swin-Transformer-Object-Detection.git

• Execute the following command ：

    cd Swin-Transformer-Object-Detection
    python setup.py develop

• Test for successful installation
  First of all to Swin-Transformer-Object-Detection Of github Official website https://github.com/SwinTransformer/Swin-Transformer-Object-Detection Download the weight file ,
  

Then put the weight file into Swin-Transformer-Object-Detection In the folder

Finally, enter the following test command on the command line , If the following result appears , The installation is successful ：

python demo/image_demo.py demo/demo.jpg 
configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_1x_coco.py 
mask_rcnn_swin_tiny_patch4_window7_1x.pth

Be careful ： The above test command is actually a line of command , The effect of entering it into the command line is as follows

If a message appears UserWarning: Matplotlib is currently using agg, which is a non-GUI backend , Can be found in demo/image_demo.py Add the following

import matplotlib
matplotlib.use('TkAgg')