Detectron2 installation based on Anaconda under win10

win10 Based on anaconda Of detectron2 install

stay Anaconda Create a virtual environment in

for example , I named the virtual environment I created mPytorch, Execute the following command to create this environment ：

conda create -n mPytorch python=3.7

After creation , Use conda activate Statement to activate the environment . Which environment is currently activated , function conda install perhaps pip install The installed library exists in which environment . Activate mPytorch after , Run the official command to install pytorch and cudatoolkit.

conda activate mPytorch
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch

install cuda Of C Language development package

Be careful , Previous section cudatoolkit Which version is installed , Go to NVIDIA's official website to download the corresponding version C Language cuda Development kit , And install . Its download and installation reference The website . Download Address .
When the download is complete , Double click the installation package , Accept the default installation options , Until installation is complete .

install MSVC compiler

If downloaded cuda10.x edition , You should install VS2019; If downloaded cuda11.x edition , You should install VS2022.
After installation , It should be found in the start menu x64 Native Tools Command Prompt for VS 2019 Command line entry ：

Download and compile the installation detectron2

Although the official claim does not support windows Download and install , But the test is feasible .

stay GitHub download detectron2 And decompress , Or use git clone Command to copy the warehouse to the local （ If you install it Git for windows Words ）.

The source code folder is shown in the figure ：

According to the official tutorial , To compile and install detectron2, Just execute a command on the command line ：

python -m pip install -e .\detectron2-main\

among -e Followed by the source folder path , It can be absolute path or relative path .
however ,windows I want to complete the compilation successfully , The following points need to be noted ：

1. Install a series of dependencies ：

pip3 install easydict
conda install pyyaml
conda install cffi
conda install cython
conda install scipy
pip install pycocotools

3. Open your x64 Native Tools Command Prompt for VS 2019 or x64 Native Tools Command Prompt for VS 2022, And enter the conda activate mPytorch Activate conda Environmental Science .
4. Set temporary environment variables

set DISTUTILS_USE_SDK=1

5. Modify the source file nms_rotated_cuda.cu:
   
   open nms_rotated_cuda.cu, Use the framed part in the following figure at the beginning of the line ’//' Just comment it out . This is to solve the error message nms_rotated_cuda.cu(68): error: identifier "single_box_iou_rotated" is undefined, You can also refer to This blog .
   
6. function python -m pip install -e .\detectron2-main\ Compile and install .

Test use detectron2

Need to install opencv.

pip install opencv-python

According to the official example , Download a trained model ,model_final_280758.pkl, And put it in detectron2 Source code demo Under the folder .
Then enter on the command line demo Folder , Run the command ：

python .\demo.py --config-file ..\configs\COCO-Detection\faster_rcnn_R_50_FPN_3x.yaml --input  Your input picture name   --opts MODEL.WEIGHTS model_final_280758.pkl