The config file includes data path, optimizer, scheduler, etc, ...

In each configure file:

• stages/data_params/root: To the folder where stores image data.
• image_size: determine the size of image

Note:

You do not need to change: train_csv and valid_csv because they are overrided by running bash file bellow.

The following data is used for different models.

• 3 windows (3w) data:

  python src/preprocessing.py extract-images --inputdir <kaggle_input_dir> --outputdir <output_folder>

• 3 windows (3w) with crop data:

  python src/preprocessing_3w.py extract-images --inputdir <kaggle_input_dir> --outputdir <output_folder>

• 3d data:

  python src/preprocessing2.py

• Start docker:

  make run
  make exec 
  cd /kaggle-rsna/

• Train resnet18, resnet34, resnet50, alexnet with 3 windows (3w) setting:

  bash bin/train_bac_3w.sh 

  Note: normalize=True

• Train resnet50 with 3d setting:

  bash bin/train_bac_3d.sh 

  Note: normalize=False

• Train densenet169 with 3 windows and crop setting:

  bash bin/train_toan.sh 
  bash bin/train_toan_resume.sh

  Note: normalize=True

where:

• CUDA_VISIBLE_DEVICES: GPUs number required to train.
• LOGDIR: Output folder which stores the checkpoints, logs, etc.
• model_name: the name of model to be trained. The script supports the name of model in here
• It is better to create a wandb account, it will help you track your log, backup the code, store the checkpoints on the could in real-time. If you dont want to use wandb, please set: WANDB=0

Output:

The best checkpoint is saved at: ${LOGDIR}/${log_name}/checkpoints/best.pth.

python src/inference.py

Check function predict_test_tta_ckp for more information, you may want to change the path, the name of model and the output path. For 3d setting, normalization=False, otherwise normalization=True

In src/ensemble.py, you should change the prediction path of each fold of model and the name of output ensemble.

python src/ensemble.py