Skip to content

Latest commit

 

History

History
80 lines (66 loc) · 2.85 KB

README.md

File metadata and controls

80 lines (66 loc) · 2.85 KB
Raw

Train NNs

Before Usage

Manually set TensorFlow backend for KERAS, add this for example to your ~/.zshrc

export KERAS_BACKEND=tensorflow

Adjust settings

  • add/change samples used in training
input_samples.addSample("SAMPLENAME"+naming, label = "SAMPLENAME", normalization_weight = FLOAT)

for example

input_samples.addSample("ttHbb"+naming, label = "ttHbb", normalization_weight = 2.)
input_samples.addSample("ttbb"+naming,  label = "ttbb")

  • change network architecture config if optimizing, otherwise add network architecture to the file net_configs.py as dictonary entry in the config_dict to not lose it and execute it with the parser option --netconfig=CONFIGNAME

  • only change DNN training class dnn = DNN.DNN([...]) properties

    • eval_metrics
    • test_percentage percentage of samples used to test

    others are changed with parser options!

Usage

To execute with default options use

python train_template.py

or use the following options

  1. Category used for training

    • -c STR name of the category (ge)[nJets]j_(ge)[nTags]t (default is 4j_ge3t)

  2. Naming/File Options

    • -o DIR to change the name of the ouput directory, can be either a string or absolute path (default is test_training)
    • -i DIR to change the name of the input directory, can be either a string or absolute path (default is InputFeatures)
    • -n STR to adjust the name of the input file generated in preprocessing (default is dnn.h5)
    • -v FILE to change the variable Selection, if the file is in /variable_sets/ the name is sufficient, else the absolute path is needed (default is example_variables)

  3. Training Options

    • -e INT change number of training epochs (default is 1000)
    • -s INT change number of epochs without decrease in validation loss before stopping (default is 20)
    • --netconfig=STR STR of the config name in net_config dictonary innet_configs.py (config in this file will not be used anymore!)

  4. Plotting Options

    • -p to create plots of the output
    • -l to create logarithmic plots
    • --printroc to print ROC value for confusion matrix
    • --privatework to create private work label
    • --signalclass=STR to change the plotted signal class (not part of the background stack plot), possible to do a combination, for example ttHbb,ttbb (default is None)

Example:

python train_template.py -i /path/to/input/files/ -o testRun --netconfig=test_config --plot --printroc -c ge6j_ge3t --epochs=1000 --signalclass=ttHbb,ttbb

Re-evaluate DNN after training with

python eval_template.py

using the option -i DIR to specify the path to the already trained network.

The plotting options of train_template.py are also avaiable for this script.

Example:

python eval_template.py -i testRun_ge6j_ge3t -o eval_testRun --plot --printroc