MEMFlow

Installation

Docker image

docker run -ti --gpus all gitlab-registry.cern.ch/dvalsecc/memflow:latest python

Singularity image

A singularity image is unpacked on the CVMFS system: /cvmfs/unpacked.cern.ch/gitlab-registry.cern.ch/dvalsecc/memflow:latest

apptainer shell --bind /afs -B /cvmfs/cms.cern.ch --bind /tmp  --bind /eos/cms/  /cvmfs/unpacked.cern.ch/gitlab-registry.cern.ch/dvalsecc/memflow:latest

Environment on lxplus-gpu

source /cvmfs/sft.cern.ch/lcg/views/LCG_103cuda/x86_64-centos9-gcc11-opt/setup.sh
python -m venv myenv
source myenv/bin/activate
git clone git@github.com:valsdav/MEMFlow.git
cd MEMFlow
pip install -e .
cd ..
git clone git@github.com:valsdav/zuko.git
cd zuko
pip install -e .

cd ..
pip install jupyterlab
jupyter lab build
# register the environment in jupyter lab
python -m ipykernel install --user --name=myenv

Matrix Element evaluation

• Start apptainer image with madgraph and LHAPDF
• Generate the process
• Compile it for python with https://cp3.irmp.ucl.ac.be/projects/madgraph/wiki/FAQ-General-4

How to run the scripts:

scripts/run_generate_config.py

• Create multiple config files. All the possible configurations will be generated (as an example iterate over learning-rate list, hidden-layers list etc.). The config file keeps: input dataset, input shape, model and training parameters.

To run it:

python scripts/run_generate_config.py --input_dataset=<path-to-Dataset> --maxFiles=-1 <--preTraining>

The argument maxFiles represent the maximum number of config files generated. By default, it is -1 (generate all possible config files). The preTraining flag makes the script iterate only on the condTransformer hyperparameters (unfolding flow parameters are set to the default value).

scripts/run_model.py

• Run the model (training and validate loops) for a specific config files.

To run it:

python scripts/run_model.py --model-dir=<path> <--on-GPU>

--model-dir: path to directory where the config file and ConditionalTransformer weights (from pretraining) are saved.

If --on-GPU flag is added, the script will run on GPU.

scripts/sendJobs.py

• Send jobs:

To run it:

python scripts/sendJobs.py --config-directory=<configDir> <--on-GPU> <--preTraining>

By default, the script is running on CPU. If --on-GPU flag is added, the script will run on GPU. The script will iterate over config-Directory and will send a job for every config file. If --preTraining flag is set, the script will send jobs for run_pretraining.py, otherwise it will send jobs for run_model.py

CAREFUL! Some paths are already set in the script, so modify them before using!!