dctr-ttbar

This module contains a modified DCTR approach to find reweighting between MadGraph EFT LO+jet sample and the central NLO Powheg sample.

Setup

If conda or micromamba are not already available, choose one and install:

curl https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh > conda-install.sh
bash conda-install.sh

Or,

"${SHELL}" <(curl -L micro.mamba.pm/install.sh)

Next, clone this repository:

git clone [email protected]:hannahbnelson/dctr-ttbar.git dctr
cd dctr

Then, setup the micromamba environment:

unset PYTHONPATH 
unset PERL5LIB
micromamba env create -f environment.yml
micromamba activate dctr
pip install -e .

The -e option installs the project in editable mode (i.e. setuptools "develop mode"). If you wish to uninstall the package, you can do so by running pip uninstall dctr. This repository also depends on the topcoffea package, which is not yet available on PyPI, so clone the topcoffea repo and install it.

cd 
git clone https://github.com/TopEFT/topcoffea.git
cd topcoffea
pip install -e .  

Now all of the dependencies have been installed and the dctr repository is ready. In the future, just activate the environment:

unset PYTHONPATH
unset PERL5LIB
micromamba activate dctr

Run a job with work_queue

Submit workers on glados

The workers must be submitted from the same environment that you are running the run script from so open a new ssh session to glados and run these commands:

unset PYTHONPATH
unset PERL5LIB
micromamba activate dctr
condor_submit_workers -M ${USER}-workqueue-coffea -t 900 --cores <Ncores> --memory <Nmemory(MB)> --disk 100000 <Nworkers>

The workers will terminate themselves after 15 minutes of inactivity.

Submit workers on the ND CRC opportunistic resources

First, login to the ND CRC condor node and activate the conda environment. Normally, the same condor_submit_wokers command from above works to submit workers. But currently, there is an authentication issue that causes an error when trying to run condor jobs on condorfe while accessing files via xrootd. Temporarily, use the script condor_custom to submit workers using the following command:

ssh glados
ssh condorfe.crc.nd.edu
unset PYTHONPATH
unset PERL5LIB
micromamba activate dctr

python condor_custom  --manager ${USER}-workqueue-coffea --cores <Ncores> --memory <Nmemory (MG)> --disk 100000 --num-workers <Nworkers> 

Scripts

• analysis/data_processor.py:
  • creates the pandas dataframes containing the variables used for neural network training
  • applies top mass cuts removing all events with top mass < 150 GeV or > 195 GeV
• analysis/run_NN_processor.py:
  • run script for analysis/data_processor.py and other coffea processors
  • can run with futures for small local tests, or work_queue for full runs
• analysis/df_accumulator:
  • accumulator class for pandas dataframes
  • used by analysis/data_processor.py
• analysis/make_pytorch_datasets.py:
  • takes pkl files containing dataframes produced by analysis/data_processor.py shuffles, and separates the events into train/validation/testing files
  • produces new pkl files containing subsets of the original ones to use as the inputs for analysis/train.py
• analysis/config.yaml
  • config file for train.py
  • needs to contain these entries at minimum: inputs, model, params, monitoring
• analysis/train.py:
  • trains neural network based on inputs from config.yaml
  • can be run locally for a quick test using command line options (--config, --outdir, --cores)
  • submit as a condor job using condor_submssions/run_train.py
• condor_submissions/run_train.py:
  • run with command line options to optionally specify the training script, config file, output directory, cores/memory per job