Vesuvius Challenge - Ink Detection

https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/overview

Environment

• Docker: 23.0.1
• rocker: 0.2.10
• off-your-rocker

Setup

Build a docker image.

make build-docker
pipx install git+https://github.com/sloretz/off-your-rocker --include-deps

Download data from the competition page.

$ tree /path_to_input -L 2
/path_to_input
└── vesuvius-challenge-ink-detection
    ├── sample_submission.csv
    ├── test
    └── train

Download pre-trained models of 3D-ResNet.

gdown 12FxrQY2hX-bINbmSrN9q2Z5zJguJhy6C
gdown 1K9oiny9ENYODFxjFBTKdGeoDvOzPu4qQ
mkdir checkpoints
mv r3d18_KM_200ep.pth checkpoints
mv r3d34_KM_200ep.pth checkpoints

Experiments

make run-docker LOCAL_DATA_PATH=/path_to_input

Preprocess

python -m inkdet.tools.preprocess

Train

train <CONFIG_PATH>

Run shell scripts in the scripts directory to train models for the submission.

Reference

• https://github.com/kenshohara/3D-ResNets-PyTorch