nhtlongcs / aic2021-thefirstswans

Evaluate

cd this-repo
python convert/label2pred.py <submission-folder-path>
python evaluation_det_e2e_offline/main.py <mode> text_results.json

Ensemble

Update by dictionary

cd this-repo
python ensemble/update_dict.py <submission-folder-path>/* --threshold 0.5 --output_dir outputs/

Tuning hyperparameters

cd this-repo
python ensemble/params_search.py <submission-folder-path> --study <save-study-name>

Docker

Pull already built docker:

docker pull vinhloiit/aic2021:latest

Or build using command (see build preparation below):

DOCKER_BUILDKIT=1 docker build --build-arg USERNAME=aic -t aic:latest .

Run command (final):

docker run --rm -it --gpus device=0 -v "$(pwd)"/TestB1:/data/test_data:ro -v "$(pwd)"/submission_output:/data/submission_output aic:latest /bin/bash run.sh

Go to inside docker container:

docker run --rm --gpus device=0 -it -v "$(pwd)"/test_data:/data/test_data -v "$(pwd)"/submission_output:/data/submission_output aic:latest /bin/bash

Build preparation

• Put images that we want to test (e.g. TestB1) to root folder, e.g ./TestB1. Or you can modify the path in the running command above (-v "$(pwd)"/TestB1:/data/test_data:ro. Format: <path outside docker>:<path inside docker>:ro, ro for read-only)
• Download weights to weights folder
• Update weight paths and run command in run.sh