arena_elo/update_elo_rating.sh

# set LOGDIR to default if not set before
if [ -z "$LOGDIR" ]; then
    echo "LOGDIR is not set. Using default '../GenAI-Arena-hf-logs/vote_log'"
    export LOGDIR="../GenAI-Arena-hf-logs/vote_log"
fi
set -e
# exit if logdir does not exist
if [ ! -d "$LOGDIR" ]; then
    echo "LOGDIR does not exist. Please check the path."
    exit 1
fi

mkdir -p results

# # for battle data
python -m elo_rating.clean_battle_data --task_name "image_editing"
edition_battle_cutoff_date=`cat cut_off_date.txt` && rm cut_off_date.txt && echo "Image editing battle data last updated on $edition_battle_cutoff_date"

python -m elo_rating.clean_battle_data --task_name "t2i_generation"
generation_battle_cutoff_date=`cat cut_off_date.txt` && rm cut_off_date.txt && echo "T2I image generation battle data last updated on $generation_battle_cutoff_date"

python -m elo_rating.clean_battle_data --task_name "video_generation"
video_generation_battle_cutoff_date=`cat cut_off_date.txt` && rm cut_off_date.txt && echo "Video generation battle data last updated on $video_generation_battle_cutoff_date"


mkdir -p ./results/$edition_battle_cutoff_date
mkdir -p ./results/$generation_battle_cutoff_date
mkdir -p ./results/$video_generation_battle_cutoff_date

cp clean_battle_image_editing_$edition_battle_cutoff_date.json ./results/latest/clean_battle_image_editing.json
cp clean_battle_t2i_generation_$generation_battle_cutoff_date.json ./results/latest/clean_battle_t2i_generation.json
cp clean_battle_video_generation_$video_generation_battle_cutoff_date.json ./results/latest/clean_battle_video_generation.json
mv clean_battle_image_editing_$edition_battle_cutoff_date.json ./results/$edition_battle_cutoff_date/clean_battle_image_editing.json
mv clean_battle_t2i_generation_$generation_battle_cutoff_date.json ./results/$generation_battle_cutoff_date/clean_battle_t2i_generation.json
mv clean_battle_video_generation_$video_generation_battle_cutoff_date.json ./results/$video_generation_battle_cutoff_date/clean_battle_video_generation.json


echo "Calculating Elo rating for image editing task"
python3 -m elo_rating.elo_analysis --clean-battle-file ./results/$edition_battle_cutoff_date/clean_battle_image_editing.json
edition_battle_cutoff_date=`cat cut_off_date.txt` && rm cut_off_date.txt && echo "Image editing battle data (after filtering models with < 50 battles) last updated on $edition_battle_cutoff_date"
mv ./elo_results_$edition_battle_cutoff_date.pkl ./results/$edition_battle_cutoff_date/elo_results_image_editing.pkl

echo "Calculating Elo rating for t2i generation task"
python3 -m elo_rating.elo_analysis --clean-battle-file ./results/$generation_battle_cutoff_date/clean_battle_t2i_generation.json
generation_battle_cutoff_date=`cat cut_off_date.txt` && rm cut_off_date.txt && echo "T2I image generation battle data (after filtering models with < 50 battles) last updated on $generation_battle_cutoff_date"
mv ./elo_results_$generation_battle_cutoff_date.pkl ./results/$generation_battle_cutoff_date/elo_results_t2i_generation.pkl

echo "Calculating Elo rating for video generation task"
python3 -m elo_rating.elo_analysis --clean-battle-file ./results/$video_generation_battle_cutoff_date/clean_battle_video_generation.json
video_generation_battle_cutoff_date=`cat cut_off_date.txt` && rm cut_off_date.txt && echo "Video generation battle data (after filtering models with < 50 battles) last updated on $video_generation_battle_cutoff_date"
mv ./elo_results_$video_generation_battle_cutoff_date.pkl ./results/$video_generation_battle_cutoff_date/elo_results_video_generation.pkl

# generat the leaderboard

python -m elo_rating.generate_leaderboard \
    --elo_rating_pkl "./results/$edition_battle_cutoff_date/elo_results_image_editing.pkl" \
    --output_csv "./results/$edition_battle_cutoff_date/image_editing_leaderboard.csv"

python -m elo_rating.generate_leaderboard \
    --elo_rating_pkl "./results/$generation_battle_cutoff_date/elo_results_t2i_generation.pkl" \
    --output_csv "./results/$generation_battle_cutoff_date/t2i_generation_leaderboard.csv"

python -m elo_rating.generate_leaderboard \
    --elo_rating_pkl "./results/$video_generation_battle_cutoff_date/elo_results_video_generation.pkl" \
    --output_csv "./results/$video_generation_battle_cutoff_date/video_generation_leaderboard.csv"

mkdir -p ./results/latest
cp ./results/$edition_battle_cutoff_date/image_editing_leaderboard.csv ./results/latest/image_editing_leaderboard.csv
cp ./results/$generation_battle_cutoff_date/t2i_generation_leaderboard.csv ./results/latest/t2i_generation_leaderboard.csv
cp ./results/$video_generation_battle_cutoff_date/video_generation_leaderboard.csv ./results/latest/video_generation_leaderboard.csv
cp ./results/$edition_battle_cutoff_date/elo_results_image_editing.pkl ./results/latest/elo_results_image_editing.pkl
cp ./results/$generation_battle_cutoff_date/elo_results_t2i_generation.pkl ./results/latest/elo_results_t2i_generation.pkl
cp ./results/$video_generation_battle_cutoff_date/elo_results_video_generation.pkl ./results/latest/elo_results_video_generation.pkl