MLPF FCC

particle-flow with ml

ML pipeline:

• The dataloaders, train scripts and tools are currently based on Weaver, the reason for this is that we are importing a root file that contains the dataset and these files can be large. Weaver has all the tools to read and load from the rootfile and also develops and iterable dataloader that prefetches some data. Currently this dataset includes events. One event is formed by hits (which can be tracks or calo hits). An input is an event in the form of a graph, and the output is a single particle (in coming versions of the dataset there will be more).
• Models: The goal of the current taks is to regress the particle's information (coordinates and energy). Currently the best approach is the [object condensation]([https://link-url-here.org](https://github.com/hqucms/weaver-core/tree/main](https://arxiv.org/abs/2002.03605), since it allows to regress a variable number of particles.
• Training: To train a model run the following command python -m src.train --data-train /eos/user/m/mgarciam/datasets/pflow/tree_mlpf2.root --data-config config_files/config_2_newlinks.yaml --network-config src/models/wrapper/example_gravnet_model.py --model-prefix models_trained/ --num-workers 0 --gpus --batch-size 100 --start-lr 1e-3 --num-epochs 1000 --optimizer ranger --fetch-step 1 --log logs/train.log --log-wandb --wandb-displayname test --wandb-projectname mlpf --condensation

Currently this model does not train because we need to remove from the dataset the events where there are no links between all of the hits and the particles (i.e all hits are noise)

Debugging the model, experimenting with (some) hyperparameters etc.

You can add parameters that get passed as kwargs to the model wrapper in the config file:

custom_model_kwargs:
   # add custom model kwargs here
   # ...
   n_postgn_dense_blocks: 4

Visualization

Runs for this project can be found in the following work space: https://wandb.ai/imdea_dolo/mlpf?workspace=user-imdea_dolo

Environment

To set up the env create a conda env following the instructions from Weaver and also install the packages in the requirements.sh script above

Alternatively, you can try to use a pre-built environment from this link - the .tar.gz file was built using conda-pack on fcc-gpu-04v2.cern.ch.

Energy correction

A bit hacky for now, but the dataset is extracted from the training dataloader by adding --save-features to the training command. The dataframes used for training are saved into the cluster_features directory. Firstly a simple neural network is trained to perform energy correction. We train two models separately for neutral and charged particles.

python notebooks/13_NNs.py --prefix /eos/user/g/gkrzmanc/2024/EC_basic_model_charged --wandb_name NN_EC_train_charged --loss default --PIDs 211,-211,2212,-2212,11 --dataset-path /eos/user/g/gkrzmanc/2024/ft_ec_saved_f_230424/cluster_features/ --batch-size 8 --corrected-energy --gnn-features-placeholders 32

python notebooks/13_NNs.py --prefix /eos/user/g/gkrzmanc/2024/EC_basic_model_neutral --wandb_name NN_EC_train_neutral --loss default --PIDs 130,2112,22 --dataset-path /eos/user/g/gkrzmanc/2024/ft_ec_saved_f_230424/cluster_features/ --batch-size 8 --corrected-energy --gnn-features-placeholders 32

The produced models are then loaded in src/models/GATr/Gatr_pf_e.py.

Evaluation / further training: --ec-model gat-concat or --ec-model dnn.