A multilayer perceptron framework for the classification of events into multiple mutually exclusive categories. It is derived from an older version of a previous multilayer perceptron framework for binary classification, developed by Max Welsch (https://github.com/welschma/NNFlow).

#Usage:

##Preprocessing the data

The loaded data is available in a structured-array format and has to be converted into numpy arrays with the preprocessing script. This script is invoked with the multi_get_branches.py file, where the following can be specified:

• The input paths for the training (Even) and validation (Odd) sets, each consisting of signal (ttH) and background (ttbarSL = ttbar semi-leptonic) events.

• The branchlists to be converted; expects a list of strings, each of them describing a file containing the branches to be converted.

• The categories to be converted; expects a list os list of strings. Each of the lists will be extracted. The code for the background events is:

  • '30': tt + bb
  • '20': tt + 2b
  • '10': tt + b
  • '01': tt + cc
  • '00': tt + light flavor

  Signal events will always be converted.

• preselection:

  • 'no': no preselection
  • 'weak': >= 6 jets, >=3 btags (medium)
  • 'strong': >= 6 jets, >=4 btags (medium)

• The output directory.

The preprocessing utility also applies weights to the events and generates one-hot labels with dimension 1 + num(categories).

##Training a neural network

The training of a neural network is invoked using the train.py or train_ttH.py file. Several parameters are to be set prior to launch:

• The paths to the training and validation sets. A path to a file containg the cross section weights for signal and background is needed in order to plot a weighted heatmap. In order to determine the most heavily weighted input features, a path to the converted branchlist must be given.

• outpath and exec_name: The models will be stored in concat(outpath, exec_name).

• Labels: The labels of the converted categories. Only needed for plotting reasons.

• normalization: The normalization algorithm to use. Can be either 'gaussian' or 'minmax'.

• outsize: The dimension of the output vector. Must be the same as len(labels).

• optname: The optimizer to choose. The following are implemented:

  • 'Adam': Adam optimizer
  • 'GradDescent': Gradient Descent optimizer
  • 'Adagrad': Adagrad optimizer
  • 'Adadelta': Adadelta optimizer
  • 'Momentum': Momentum optimizer

• optimizer_options: The options for the chosen optimizer. Refer to the train.py and train_ttH.py files and to the perceptron framework for details of possible values. If given an empty list, the default options are used.

• act_func: The activation function. The following values are supported:

  • 'tanh'
  • 'relu'
  • 'elu'
  • 'softplus'
  • 'sigmoid

  If no activation function is chosen, or another value is passed, 'tanh' will be chosen.

• beta: The L2 regularization parameter.

• N_EPOCHS: Number of epochs to be trained if no early stopping occurs.

• learning_rate: Optimizer learning rate.

• batch_size: Size of the training batch.

• hidden_layers: A list containing the number of neurons in each layer. The list [200, 100, 100] will produce a neural network with three hidden layers, of which the first contains 200 neurons, while the latter two letters each contain 100 neurons.

• keep_prob: The firing probability for a neuron. Must be a value 0 < keep_prob <= 1.

• decay_learning_rate: Indicates whether to manually decay the learning rate. Allowed values are 'yes' and 'no'.

• lrate_decay_options: A list containing the decay rate and the step width, e.g. [0.97, 200]

• batch_decay: Indicates whether to change the batch size during the training process. Allowed values are 'yes' and 'no'.

• batch_decay_options: Sets the decay rate and step width for the batch size decay. The step_width is given in number of epochs.

• enable_early: Indicates whether to check for early stopping. Allowed values are 'yes' and 'no'. Default is 'no'.

• early_stop: The number of epochs after which the quality must have improved, otherwise early stopping is invoked.

• ttH_penalty: A penalty value for false positives and false negatives when using the ttH classifier. Not usable with train.py.

With these options, the neural network is built and trained. During the training process, plots are generated and stored.