This documentation provides the steps of training, inferencing and blending to reproduce our winning solution.
Install required dependency libraries first from the root project folder.
pip install -r requirements.txt
It is recommended to run training and inference on a local machine with at least 4 CPUs, 16GB Host Memory and 11GB GPU Memory (Nvidia 2080-Ti or better GPU).
Run the following shell script with the right absolute paths for input dataset and model artifacts. All scripts should be ran from the final folder path.
sh train.sh <input folder> <model folder>
All models will be saved under the final folder with fixed folder names.
Next, run the following shell script with the right absolute path for input dataset and trained model artifacts.
sh inference.sh <input folder> <model folder> <output folder>
All model predictions will be saved under the target output folder with fixed submission filenames.
Finally, run the following python script to create the winning solution prediction file.
python blend.py <input folder> <prediction file folder> <output folder>
That's it!