GenAI-Arena / arena_elo /update_elo_rating.sh
DongfuJiang's picture
update
f6608c4
raw
history blame
No virus
2.85 kB
# set LOGDIR to default if not set before
if [ -z "$LOGDIR" ]; then
echo "LOGDIR is not set. Please set LOGDIR to the directory where the logs will be stored. Command: export LOGDIR=/path/to/logdir"
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 $generation_battle_cutoff_date"
mkdir -p ./results/$edition_battle_cutoff_date
mkdir -p ./results/$generation_battle_cutoff_date
cp clean_battle_image_editing_$edition_battle_cutoff_date.json ./results/latest/
cp clean_battle_t2i_generation_$generation_battle_cutoff_date.json ./results/latest/
mv clean_battle_image_editing_$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/$generation_battle_cutoff_date/clean_battle_t2i_generation.json
python3 -m elo_rating.elo_analysis --clean-battle-file ./results/$edition_battle_cutoff_date/clean_battle_image_editing.json
mv ./elo_results_$edition_battle_cutoff_date.pkl ./results/$edition_battle_cutoff_date/elo_results_image_editing.pkl
python3 -m elo_rating.elo_analysis --clean-battle-file ./results/$generation_battle_cutoff_date/clean_battle_t2i_generation.json
mv ./elo_results_$generation_battle_cutoff_date.pkl ./results/$generation_battle_cutoff_date/elo_results_t2i_generation.pkl
# generat the leaderboard
python -m elo_rating.generate_leaderboard \
--model_info_file "./edition_model_info.json" \
--elo_rating_pkl "./results/$edition_battle_cutoff_date/elo_results_image_editing.pkl" \
--output_csv "./results/$edition_battle_cutoff_date/image_editing_leaderboard.csv"
python -m elo_rating.generate_leaderboard \
--model_info_file "./generation_model_info.json" \
--elo_rating_pkl "./results/$generation_battle_cutoff_date/elo_results_t2i_generation.pkl" \
--output_csv "./results/$generation_battle_cutoff_date/t2i_generation_leaderboard.csv"
mkdir -p ./results/latest
cp ./results/$edition_battle_cutoff_date/image_editing_leaderboard.csv ./results/latest/image_editing_leaderboard.csv
cp ./results/$generation_battle_cutoff_date/t2i_generation_leaderboard.csv ./results/latest/t2i_generation_leaderboard.csv
cp ./results/$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