New data of Hss plus evaluation commands - selvaggi/mlpf GitHub Wiki

create data

cd mlpf/condor_CLD
myschedd bump # to get a better condor alloc % changed the paths of outdir and condor dir of script_create_dataset_eval and changed the mlpf/condor_CLD/run_sequence_CLD_eval.sh to Zcard
bash script_create_dataset_eval.sh

(created 400 files of 10 events each)

evaluate on this data:

If you want to change to a certain clustering pretrained model change the mlpf/src/utils/load_pretrained_models.py in the load_test_model function, because the --load-model-weights in the arguments first loads a model that includes the energy correction trained on ground truth clusters and changing the clustering just overwrites the clustering part.
You can also change the clustering parameters in mlpf/src/layers/inference_oc.py in function hfdb_obtain_labels.
python -m src.train_lightning1 --data-test /eos/experiment/fcc/users/m/mgarciam/mlpf/CLD/eval/2_particles/Hss_250525/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_Hss_comp/ --wandb-displayname eval_gun_drlog --num-workers 0 --gpus 0 --batch-size 10 --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

The results of evaluating this data are here /eos/user/m/mgarciam/datasets_mlpf/models_trained_CLD/eval_comp_Hss/gun_drlog_v9_99500_hbdscan_Hss_400_15_12_01/ the efficiency and fake rate seem to be pretty similar to what we had in the note. The mass res is 0.0977 (worse than the Hss training and worse than pandora)

Part 2. Train with physics events and evaluate again vs training with gun

The training that was used for the note is:
https://wandb.ai/ml4hep/mlpf_debug/runs/imme1iwf/overview
The eval script:
https://wandb.ai/fcc_ml/mlpf_debug_eval/runs/l9x75ec9