train_tutorial.md

Training tutorial

• This tutorial can walk through how to load pretrained model and run the training script.

environment

conda create --name py36-torch python=3.6
pip install -r requirements.txt

Required package

• roi_pool:
  • install the package: https://github.com/open-mmlab/mmdetection/blob/master/INSTALL.md
  • https://github.com/open-mmlab/mmdetection/tree/master/mmdet/ops/roi_pool
  • put 'roi_pool_cuda.cpython-36m-x86_64-linux-gnu.so' in 'utils/roi_pool/'

Required settings

• check the config file
  • check the model path
    • set in 'pretrained'
    • set 'retrain' to false
    • (set 'reset_iter' to false)
  • check the data path: (can use hyperlink: ln -s)
    • put data in the path 'dataset'
      • kitti: 'datasets/kitti_wVal'(default in setting.py)
    • the folder name should match the one listed on ['data']['dataset']
    • put the files in 'datasets/kitti_split' to 'datasets/kitti_wVal'

     cp datasets/kitti_split/train.txt datasets/kitti_wVal/
     cp datasets/kitti_split/val.txt datasets/kitti_wVal/
    

  • check the labels path
    • check the path: ['data']['labels']
      • kitti: logs/magicpoint_synth20_homoAdapt100_kitti_h384/predictions (default)
    • the path uses base path EXPER_PATH (listed in settings.py)

Run the code

python train4.py <train task> <config file> <export folder>
python train4.py train_joint --debug configs/superpoint_kitti_train_heatmap.yaml superpoint_kitti --eval

Related files

• train4.py: training script (load 'train_model_frontend')
• train_model_frontend.py: class for training
• configs/superpoint_coco_train.yaml: path and parameter settings

Code logic

testing log

• 2019/7/11
  • test python train4.py train_joint --debug configs/superpoint_kitti_train_heatmap.yaml superpoint_kitti --eval
  • environment: python: 3.6, pytorch: 1.1, cuda:10