Dump an H5 File

Overview

Teaching: 10 min
Exercises: 10 min

Questions

How do I get data to Keras

Objectives

Get a data file you can use

Get your credentials etc

The script in the analysisbase-launcher expects that your key is in ~/.ssh/id_rsa. If you don’t want to use this you can run make-key.sh and copy the output into your gitlab SSH keys.

Set up your working directory

You’ll need to make your input data available within the image. Copy it into the work/data directory.

Building the dumper

You can load up the image by running ./run-docker.sh script. This should (hopefully) drop you into a working image. I’ve added most of the annoying setup stuff to the .bashrc so you should be ready to clone and build.

Clone and build the dumper with

ssh://git@gitlab.cern.ch:7999/deep-sets-example/higgs-regression-dumper.git

Then make a build directory and build it.

Producing an output file

You’ll have to run something like

./AnalysisPayload ../data/DAOD_FTAG5.p3870.pool.root

Your path may be something else.

This will produce an output file called jets.h5. Let’s take a look at it:

h5ls -v jet.h5

This prints a lot of information:

Opened "jets.h5" with sec2 driver.
jets                     Dataset {18757/Inf}
    Location:  1:1552
    Links:     1
    Chunks:    {2048} 24576 bytes
    Storage:   225084 logical bytes, 201354 allocated bytes, 111.79% utilization
    Filter-0:  deflate-1 OPT {7}
    Type:      struct {
                   "GhostHBosonsPt"   +0    native float
                   "GhostBHadronsFinalPt" +4    native float
                   "pt"               +8    native float
               } 12 bytes
subjets                  Dataset {18757/Inf, 5/5}
    Location:  1:800
    Links:     1
    Chunks:    {2048, 5} 122880 bytes
    Storage:   1125420 logical bytes, 517620 allocated bytes, 217.42% utilization
    Filter-0:  deflate-1 OPT {7}
    Type:      struct {
                   "pt"               +0    native float
                   "rnnip_pb"         +4    native float
                   "JetFitter_mass"   +8    native float
               } 12 bytes

We’ll be trying to learn GhostHBosonPt from the other variables. Note that subjets is a 2d array, we’re saving all the subjets in each large-R higgs jet.

Key Points

We always reformat our data before M’Learnin

lesson home

Machine Learning

next episode