M'learn some higgs regression

Overview

Teaching: 10 min
Exercises: 10 min

Questions

• How do I train NNs

Objectives

• Train a NN for higgs regression

Setting up your environment

I like to run these simple tests on my laptop natively, but if for whatever reason you can’t do that, you can run all this within an image.

We’d recommend you run

docker pull atlasml/ml-base:py-3.7.2

to get an image to start with

What are we trying to do?

The idea here is to regress on the true higgs pt based on some information that we’re storing in the jets. Whether this makes any sense is something to talk about with physicists, but we just want to demonstrate the workflow here.

Get the training / plotting repo

The repository is called higgs-regression-training. Clone it in the /work directory.

Launching into another image

If you need a docker image to run python 3, Keras, etc, you can get it by going to the work directory and running.

docker run --rm -it -v ${PWD}:/home/atlas/data atlasml/ml-base:py-3.7.2

Making a plot

Try running draw_regression.py -h This should prompt you for some arguments like the H5 file you just produced. Once you’ve fed in the right augments it should produce a plot in plots/ by default. What we want to do is make this line more diagonal.

Running the training

Now try running train_nn.py -h. If you figure out the right arguments and feed them to the training script, you should see a lot of status bars that mean the NN is training. Great! The model will be saved in model.

After you run the training, you should see a few files in a model directory. Run tree model:

model/
├── architecture.json
├── model.png
└── weights.h5

There are a few files:

• The architecture.json shows the general structure of the model, but this is Keras’s own special language
• The elements of the dense layers are stored in the weights.h5
• The more “human readable” format is in model.png

Take a look at model.png and try to make sense of it.

Key Points

• This part should be pretty easy