Kyle notified people over email and Twitter that Valentin Hirschi had gotten MadGraph to export matrix element codes for Feynman diagrams into a form that can be used by differentiable machine learning libraries (e.g., TensorFlow and PyTorch).
The project is called MG5aMC_PythonMEs.
This is a short walkthrough of doing the simplest possible thing: just running the code that MG5aMC_PythonMEs generates.
As I don't have much experience with MadGraph5_aMC@NLO and its installation and use instructions are quite bad I threw together a Docker image that at leasts somewhat works. This walkthrough will use that.
To be clear, this is never how you would actually do anything with a Docker image. This is just throwing things against the wall at the moment.
Get the Docker image I made from the Docker Hub image registry
docker pull matthewfeickert/madgraph5-amc-nlo
I'm assuming you're in the directory of this Gist, so just volume mount locally
docker run --rm -it -v $PWD:/home/docker/data matthewfeickert/madgraph5-amc-nlo:latest /bin/bash
then inside the container navigate to the PLUGIN directory of the MG5_aMC_v2_6_6 installation and clone the repo from GitHub (here using HTTPS instead of SSH as I don't want to bother getting SSH agent forwarding for this)
cd ~/MG5_aMC_v2_6_6/PLUGIN/
git clone https://github.com/ValentinHirschi/MG5aMC_PythonMEs.git
A missing required dependency is NumPy, and while in this very early stage example Tensorflow isn't actually required, as Valentin notes:
TFwhich is for now identical in behaviour toPythonbut uses daughter classes and templates ready to be specialised for TensorFlow (see README.md)
we'll install Python 2.7 compatible versions of both using the requirements.txt in the Gist that we have access to through the volume mount as this is an application and not a library
cd ..
pip install --user -r ~/data/requirements.txt
Now from the top level directory of the MG5_aMC_v2_6_6 installation we can run the examples given in the README of Valentin's repo.
First generating the Python 2.7 example
./bin/mg5_aMC --mode=MG5aMC_PythonMEs PLUGIN/MG5aMC_PythonMEs/test_MG5aMC_PythonMEs.mg5
and then running it
python2 MG5aMC_PythonMEs_output_example/check_sa.py
and then moving it to the local volume for a later reference
cp -r MG5aMC_PythonMEs_output_example ~/data/
Then running the TensorFlow + Python 2.7 example (which again, isn't different at the moment)
./bin/mg5_aMC --mode=MG5aMC_PythonMEs PLUGIN/MG5aMC_PythonMEs/test_MG5aMC_TFMEs.mg5
python2 MG5aMC_TFMEs_output_example/check_sa.py
cp -r MG5aMC_TFMEs_output_example ~/data/
If you were to now exit the container you would still have the two directories with scripts locally that you can poke at with your favorite editor. These are now truly standalone modules it seems, as you can exit the container and then run the Python 2.7 release container
docker run --rm -it -v $PWD:/home -w /home python:2.7 /bin/bash
get the minimal dependencies again (we'll just skip to only getting NumPy for the sake of time)
pip install --user numpy
and then you're able to run the scripts again
python2 MG5aMC_PythonMEs_output_example/check_sa.py
python2 MG5aMC_TFMEs_output_example/check_sa.py