1. Instructions to run scripts for 1D histograms and ML
Presented by Kaifu Lam
May 3, 2017

2. Background Previously presented on Mar 1 (single b-tagging)…
To re-plot the variables and understand the variables in
Jet Flavor Classifcation in High-Energy Physics with Deep Neural Networks – Sep 2016
The dataset is created by simulation modeling light and heavy jets of pp collisions in ATLAS detector.
The dataset contains expert (high) level variables and mid – low level variables:
Expert: variables (16)
Mid / low: variables X 15 tracks (422)
Variable explanations here 2

3. Expert variables and ROC

4. Python on Tev clusters
Todd’s Python installation has everything you need
Python 2.7
Python packages required: numpy, h5py, theano, keras, matplotlib
Todd’s python installation: ~olsont/TEV/python/bin/python
Tips:
Go to home ($ cd ~)
Open .cshrc file ($ nano .cshrc)
In this document, type
“alias pythonnew ~olsont/TEV/python/bin/python”
Type ‘control ‘+ ‘X’, then ‘Y’ in MacBook to exit and save
Restart tev01
Now you can run Todd’s python by typing pythonnew
Feel free to change the alias 4

5. Text Editor Suggestion
If you use Mac…
Sam Meehan recommends using TextWrangler in your MacBook
Latest version rebranded as BBEdit
Download to your personal MacBook
Use the “Open from FTP/SFTP Server” and “Save to FTP/SFTP Server” to open and save files on TevCluster, while editing locally in your MacBook
TextWrangler has saved me so much time!! 5

6. Script repositories and raw data locations
Scripts:
Single b-tagging (1D variables and ML)
H->bb tagging (1D calorimeter variables)
Raw Data:
Single b-tagging:
Tev01: /phys/groups/tev/scratch4/users/kaifulam/dguest/gjj-pheno/v1/dataset.json.gz
H->bb tagging:
Tev01: /phys/groups/tev/scratch4/users/kaifulam/dguest/hbb/v1/
Files: signal.txt, background.txt 6

7. Single b-tagging: Data transformation
Step 1: Transform raw data with Julian’s script:
Script here
Transformed data (DON’T OPEN THIS FOLDER):
/phys/groups/tev/scratch4/users/kaifulam/dguest/gjj-pheno/v1/julian/raw_data/saved_batches_test/
30 million .npy files in this folder
Example files of 10 jets are located here:
/phys/groups/tev/scratch4/users/kaifulam/dguest/gjj-pheno/v1/julian/raw_data/saved_batches/
Data structures:
3 files for each jet (using first jet as example):
High variables [1x16]: clean_diget_high_0.npy
Mid variables [1x422]: clean_dijet_mid_0.npy
Flavor [1]: clean_dijet_y_0.npy 7

8. Single b-tagging: Data transformation
Step 2: Transfer .npy files to one .h5 file
Script here
.h5 file location:
/phys/groups/tev/scratch4/users/kaifulam/dguest/gjj-pheno/v1/julian/ gjj_Variables.hdf5
Transferred only high variables and flavor variable to .h5
High_input [10m x 1 x 16]
Y_input [10m x 2] (logical array: col1 signal, col2 bg)
Shortens data reading time from 26 hours to 30 seconds
Future To-do: transfer mid level variables to .h5
[2] variables
High_input dimensions
[1] tracks
[0] Jets 8

9. Single b-tagging: Data transformation
Completed (don’t have to re-do these):
Step 1 for all variables
Step 2 for high variables and flavor variable
To be completed:
Step 2 for mid variables
Not enough RAM in tev01
Cannot add all mid variables in one take
Have to add mid variables into .h5 by batches.
histograms
1D histograms
ROC
Raw Data
Julian Export
H5 format
ML training
ROC 9

10. Single b-tagging: Scripts
*collector.csv dimensions
[1] counts per bin or bin edges
B-tagging variables
Expert level 1D histogram
Input: gjj_variables.hdf5
Output: histo_sig_collector.csv, histo_bg_collector.csv, bin_collector.csv
Plot 1D histogram (solo)
Input: histo_sig_collector.csv, histo_bg_collector.csv, bin_collector.csv
Output: One .png for each variable
Plot 1D ROC curves for each variable (One Figure)
Output: One .png for all ROC curves
Machine Learning
Expert level ML (GRU Neural network; multivariate classifier)
Output: tpr.csv, fpr.csv
Plot ROC curve for ML
Input: tpr.csv, fpr.csv
Output: One .png for ROC
[0] Variables
tpr / fpr.csv dimensions
[0] Size = bin count 10
For plotting, download (scp) input files to and run plot scripts in your local machine

11. Single b-tagging: ML Configurations
Output Layer
Current
Future
GRU layer …
[2] variables
Input layer …
High_input dimensions
[1] tracks
[0] Jets 11

12. H->bb Tagging: Data parsing
Data structure:
My Readme
Dan Guest’s Readme
Completed (don’t have to re-do these):
Parsing Raw ‘cluster’ (calorimeter) variables
cluster_pt, cluster_eta, cluster_dphi_jet, cluster_energy
Plotted variables of max energy cluster and sum of 3 max energy clusters
To be completed:
Parsing all other variables
Storing all variables in .h5
ML for all variables 12

13. In Work H->bb tagging: Scripts B-tagging variables
*collector.csv dimensions
[1] counts per bin or bin edges
B-tagging variables
1D histogram
Input: signal.txt, background.txt
Output: histo_sig_collector.csv, histo_bg_collector.csv, bin_collector.csv
Plot 1D histogram (solo)
Input: histo_sig_collector.csv, histo_bg_collector.csv, bin_collector.csv
Output: One .png for each variable
[0] Variables
tpr / fpr.csv dimensions
[0] Size = bin count 13

14. Next Steps Single b-tagging: H->bb tagging:
Store mid variables in .h5 format
Get GRU network config from Julian
Run ML using mid variables
Run ML using expert + mid variables
H->bb tagging:
Parse variables (except cluster variables)
Run ML using cluster variables
Run ML using track variables
Run ML using a combinations of all variables 14

15. Special Thanks!
Dr. Sam Meehan
Prof. Shih-Chieh Hsu 15