Detectron2 installation and testing

Just touched detectron2 frame , In general, the framework consists of Facebook Open source , The quality is very good , Worth learning . Today, I will sort out and summarize some problems encountered in installation and testing . It will be continuously updated later .

Detectron2 install

Detectron2 Project links ：https://github.com/facebookresearch/detectron2
Official installation tutorial ：https://github.com/facebookresearch/detectron2/blob/main/INSTALL.md

Requirements

Python ≥ 3.6 （ This experiment Ubuntu16.04 Proceed under , Use Python3.7）
Pytorch ≥ 1.5 Corresponding version torchvision（ This experiment is installed Pytorch 1.6,torchvision 0.7）
CUDA 10.1 （ Please choose with Pytorch Matching cudatoolkit edition , In order to successfully install the framework ）
GCC >= 4.9（ This is C/C++ Requirements of compilation environment , If you don't , In the compilation process, it will basically fail ）

Environmental installation

anaconda Create an environment

conda create -n detectron2 python=3.7
conda activate detectron2

install pytorch 1.6 cuda10.1

conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.1 -c pytorch

# Test it 
import torch, torchvision
print(torch.__version__, torch.cuda.is_available()) 

Install other dependencies

pip install cython pyyaml==5.1
pip install opencv-python==3.4.2
pip install cython
# fvcore  The following version is installed in this experiment , You can also refer to the official installation tutorial to install other versions ：https://github.com/facebookresearch/fvcore/
pip install fvcore==0.1.1.post20200716
#  install pycocotools, If the following code runs with an error , You can refer to the link for installation ：https://blog.csdn.net/xiongzai2016/article/details/106855544
pip install -U 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'

install detectron2

#  Install from local clone 
git clone https://github.com/facebookresearch/detectron2.git
python -m pip install -e detectron2

Download like this detectron2 It's the latest version , As required in the official installation documents PyTorch ≥ 1.8 , Therefore, an error will be reported during installation . You can visit detectron2 Historical version link of ：https://github.com/facebookresearch/detectron2/releases
According to the installed PyTorch and cuda Version to install the corresponding version detectron2, Here is PyTorch1.6 The corresponding installation command line ：

python -m pip install detectron2==0.4 -f \
  https://dl.fbaipublicfiles.com/detectron2/wheels/cu101/torch1.6/index.html

test - Instance segmentation

stay detectron2 Create a new folder testimage Folder , Save several test pictures .

function demo.py

python demo/demo.py \
--config-file ./configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml \
--input ./testimage/2007_003020.jpg --output ./output/ \
--opts MODEL.WEIGHTS ./model/model_final_f10217.pkl

config-file ：configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml This is the configuration information ,detectron2 There is already in that folder .
input： ./testimage/2007_003020.jpg Test the path of the image .
output： Define an output path by yourself .
opts ：MODEL.WEIGHTS models/model_final_f10217.pkl This is a trained model , Go according to your own requirements model zoom Download it . It can also be set directly to MODEL.WEIGHTS detectron2://COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/model_final_f10217.pkl Sometimes the download speed is very slow when you have time , Affect the test speed , So it's better to download it in advance .

Look at the test results ：

Training test

Use the self-contained small version dataset , function ./datasets/prepare_for_tests.sh, And then execute ./dev/run_instant_tests.sh, If the following result appears , Congratulations on the successful installation . The following is a screenshot of the running results of other bloggers ：

Reference article ：
Link to the original text ：https://blog.csdn.net/weixin_41298484/article/details/119914362
Link to the original text ：https://blog.csdn.net/XX_123_1_RJ/article/details/103779787
Link to the original text ：https://blog.csdn.net/qq_33047753/article/details/107022768