Gray Zone Assessment

Get started

1. Clone github repository

git clone https://github.com/andreanne-lemay/gray_zone_assessment.git

2. Build docker image

docker build -t gray_zone docker/

3. Run docker container

docker run -it -v tunnel/to/local/folder:/tunnel --gpus 0 gray_zone:latest bash

4. Run the following command at the root of the repository to install the modules

cd path/to/gray_zone_assessment

pip install -e .

5. Train model

python run_model.py -o <outpath/path> -p <resources/training_configs/config.json> -d <image/data/path> -c <path/csv/file.csv>

For more information on the different flags: python run_model.py --help

Configuration file (flag -p or --param-path)

The configuration file is a json file containing the main training parameters.
Some json file examples are located in gray_zone/resources/training_configs/

Required configuration parameters

csv file (flag -c or --csv-path)

The provided csv file contains the filename of the images used for training, GT labels (int from 0-n_class), patient ID (str) and split column (containing 'train', 'val' or 'test') (optional).

Example of csv file with the default column names. If the column names are different from the default values, the flags --label-colname, --image-colname, --patient-colname, and --split-colname can be used to indicate the custom column names. There can be more columns in the csv file. All this metadata will be included in predictions.csv and split_df.csv.

Output directory (flag -o or --output-path)

└── output directory                # Output directory specified with `-o`  
    ├──   checkpoints               # All models (one .pth per epoch)  
    |     ├──  checkpoint0.pth   
    |     ├──  ...  
    |     └──  checkpointn.pth   
    ├──   best_metric_model.pth     # Best model based on validation metric  
    ├──   params.json               # Parameters used for training (configuration file)  
    ├──   predictions.csv           # Test results  
    ├──   split_df.csv              # csv file containing image filenames, labels, split and patient id  
    └──   train_record.json         # Record of CLI used to train and other info for reproducibility