How to train an object detection model with mmdetection | DLology blog

Quick start

Train an object detection with Google Colab and free GPU.

Train with custom Pascal VOC dataset.

Train with custom COCO dataset.

The data/VOC2007 folder provides a reference structure of custom dataset ready for training. Fork my repository and replace them with your custom annotated dataset as necessary.

Further instruction on how to create your own datasets, read the tutorials

• How to train an object detection model with mmdetection - Custom Pascal VOC dataset.
• How to create custom COCO data set for object detection - Custom COCO dataset.

Note: Dilation,Smudge,generateVOC2JSON.py in DataPreparation are from CascadeTabNet

conda env setting

가상환경 생성

    conda create -n envName python=3.7 -y   
    conda activate envName  

이 github 프로젝트를 다운로드(sub module도 같이 다운)

    git clone --recurse-submodules $[Git-address]  

mmdetection dependency 설치

    cd mmdetection  
    pip install -q mmcv terminaltables  
    pip install -r requirements/build.txt  
    pip install torch==1.4.0+cu100 torchvision==0.5.0+cu100 -f https://download.pytorch.org/whl/torch_stable.html  

conda env에 nvcc 설치를 위한 cudatoolkit 설치

    pip install "git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI"  
    conda config --add channels conda-forge  
    conda install cudatoolkit-dev  

setup.py 내용 변경

    torch.utils.cpp_extension.CUDA_HOME = '/userhome/anaconda3/envs/cctab/pkgs/cuda-toolkit' 추가  

다시 mmdetection dependency 설치

    python setup.py install  
    python setup.py develop  
    pip install -r requirements.txt  
    pip install pillow==6.2.1  
    pip install mmcv==0.4.3  

Folder setting

Project  
├── data  
│   └── VOC2007  
│       ├── Annotations  
│       │   └── *.xml  
│       ├── JPEGImages  
│       │   └── *.jpg/png  
│       └── ImageSets  
│           └── Main  
│               ├── trainval.json  
│               └── test.json  
└── Model  
    └── cascade_mask_rcnn_hrnetv2p_w32_20e_eunji  
        └── epoch_14.pth  

data: VOC data to COCO data, by running DataPreparation/generateVOC2JSON.py
Model: set TableBank Latex (TD) pretrained model as pretrained model from CascadeTabNet

1. Locate images and VOC labels in data/VOC2007/Annotations and JPEGImages
2. make trainval/test lists

    cd ./DataPreparation  
    sh ./generate_traintest.sh  
    python divide_traintest.py --rate=0.9(train number portion)  

3. convert VOC labels to COCO label

    python generateVOC2JSON.py  

Train model

    python ./mmdetection/tools/train.py Config/cascade_mask_rcnn_hrnetv2p_w32_20e.py --gpu-ids=$GPU_NUM

Test model (detection)

    python test.py --data_dir=./data/VOC2007 --config_dir=./Config --epoch=$FINAL_EPOCH --gpu=$GPU_NUM