Clustering step configurations - selvaggi/mlpf GitHub Wiki
I explored multiple configs of the hdbscan algorithm for model /eos/user/m/mgarciam/datasets_mlpf/models_trained_CLD/gun_drlog_v9/_epoch=3_step=99500.ckpt (trained on 800k events dr log between 0.25 and 0.5):
min_cluster_size=8, min_samples=8, cluster_selection_epsilon=0.1: the results of this config are (/eos/user/m/mgarciam/datasets_mlpf/models_trained_CLD/eval_comp_05/gun_drlog_v9_99500_hbdscan__v3_130225_v1_400) which show that there is a lower efficiency on higher energy particles and higher eff of low E particles, mass resolution0.0519min_cluster_size=8, min_samples=8, cluster_selection_epsilon=0.05: results are (/eos/user/m/mgarciam/datasets_mlpf/models_trained_CLD/eval_comp_05/gun_drlog_v9_99500_hbdscan__v3_130225_v1_400_88_005) efficiency is high overall (the 0.05 fixes the low eff at higher energies but created a lot more fakes for charged and neutrals). Mass resolution0.0519min_cluster_size=15, min_samples=10, cluster_selection_epsilon=0.05: results are (/eos/user/m/mgarciam/datasets_mlpf/models_trained_CLD/eval_comp_05/gun_drlog_v9_99500_hbdscan__v3_130225_v1_400_15_12_005/), overall very similar performance in efficiency to pandora with slightly less fakes (only more fakes for charged, which probably has to do with the secondary tracks that pandora discard and we use anyways). Mass resolution0.0529.
Overall, these different clustering parameters result in different working points for the algorithm for the per-particle reconstruction while still resulting in very similar mass resolutions.
Eval command to run these:
python -m src.train_lightning1 --data-test /eos/experiment/fcc/users/m/mgarciam/mlpf/CLD/eval/gun_05_130225_v1/pf_tree_{1..400}.root --data-config config_files/config_hits_track_v4.yaml -clust -clust_dim 3 --network-config src/models/wrapper/example_mode_gatr_e.py --model-prefix /eos/user/m/mgarciam/datasets_mlpf/models_trained_CLD/eval_comp_05/ --wandb-displayname eval_gun_drlog --num-workers 0 --gpus 0 --batch-size 2 --start-lr 1e-3 --num-epochs 100 --optimizer ranger --fetch-step 0.1 --condensation --log-wandb --wandb-projectname mlpf_debug_eval --wandb-entity fcc_ml --frac_cluster_loss 0 --qmin 1 --use-average-cc-pos 0.99 --lr-scheduler reduceplateau --tracks --correction --ec-model gatr-neutrals --regress-pos --add-track-chis --load-model-weights /eos/user/g/gkrzmanc/results/2024/E_PID_02122024_dr05_s6500_3layer_pid_GTClusters_all_classes_PID/_epoch=0_step=4500.ckpt --freeze-clustering --predict --regress-unit-p --PID-4-class --n-layers-PID-head 3 --separate-PID-GATr