Record of problems in the construction process of IOD and detectron2

The paper

Incremental Object Detection via Meta-Learning (TPAMI 2021)
Incremental target detection based on meta learning

paper: https://arxiv.org/abs/2003.08798
code: https://github.com/JosephKJ/iOD

• be based on Faster R-CNN Built incremental target detector
• frame ：Detectron2
  

Code

One 、 Set up process

conda create -n iod python=3.7 -y
conda activate iod
pip install torch==1.8.1+cu101 torchvision==0.9.1+cu101 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
pip install opencv-python
pip install fvcore
pip install cython; pip install 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'

install pytorch According to cuda Choose the appropriate version https://pytorch.org/get-started/previous-versions/

I don't know if this step is necessary , But I did

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

download iod, And compile

git clone https://github.com/JosephKJ/iOD.git
cd iOD
pip install -v -e .

Two 、 Have a problem

1.error: ‘AT_CHECK’ was not declared in this scope

Report errors RuntimeError: Error compiling objects for extension, Look forward to the cause of the error , Many documents appear error: ‘AT_CHECK’ was not declared in this scope

The reason for the error ：AT_CHECK is deprecated in torch 1.5
The high version of the pytorch No longer use AT_CHECK, But use TORCH_CHECK.
resolvent ： Find the wrong file , The inside ‘AT_CHECK’ Replace all with ‘TORCH_CHECK’.

Reference resources ：https://blog.csdn.net/qq_21388689/article/details/117129404

2.linux How to start .sh file ,Linux Here's how to run .sh file

Reference resources ： https://blog.csdn.net/weixin_32149339/article/details/116583758

3.AssertionError: Checkpoint detectron2://ImageNetPretrained/MSRA/R-50.pkl not found!

download R-50.pkl To the folder , take iOD/detectron2/engine/defaults.py in 312 Change the line loading path to self download pkl Path to file

https://github.com/Majiker/BalancedMetaSoftmax-InstanceSeg/issues/3

There's another way of saying that force Version of the problem , Run it again

pip install fvcore==0.1.1.dev200512

For more solutions, please refer to ：

https://github.com/Majiker/BalancedMetaSoftmax-InstanceSeg/issues/3

4.RuntimeError: CUDA out of memory. Tried to allocate 1.53 GiB

nvidia-smi, Will be displayed GPU Usage situation , And occupy GPU Applications for
Input taskkill -PID Process number -F End the process of occupation , such as taskkill -PID 7392 -F
Input again nvidia-smi see GPU usage , Will find GPU The occupied space is greatly reduced , In this way, we can use GPU Run the program

My problem solving ：
Before typing –num-gpus 1 Use one gpu
In the service area 8 individual gpu Not used ,

5.RuntimeError: Address already in use

Port occupied

Reference resources many GPU Training appears RUNTIMEERROR: ADDRESS ALREADY IN USE solve https://www.freesion.com/article/77681373376/

CUDA_VISIBLE_DEVICES='0,1,2,3' python tools/train_net.py --dist-url tcp://127.0.0.1:50001   --num-gpus 4 --config-file ./configs/PascalVOC-Detection/iOD/base_19.yaml SOLVER.IMS_PER_BATCH 8 SOLVER.BASE_LR 0.005

a key ： --dist-url tcp://127.0.0.1:50001 Should be placed in num-gpus front , Putting it at the back has never been successful , Or report the original error , Just put it in front

Or turn off the occupied port number

Release port , Three steps are needed :

• Find all current ports of the system netstat -tln
• Find the process of the corresponding port in the system ID(PID) lsof -i : Port number
• Use kill -9 [PID] Command end process kill -9 [PID]

Reference resources Linux Release the occupied port https://cloud.tencent.com/developer/article/1340493

3、 ... and 、pycharm Connect to remote server

Select the corresponding virtual environment and the corresponding file path