Getting Started¶

How it works¶

The Borg has routines to train Artificial Neural Networks (ANNs). The two types currently supported are classifier and emulator. The models and training recipes are based on software written by Yuan-Sen Ting: The Payne for the emulator and the MWM Classifier for the convolutional classifier.

The Borg Model Basics¶

Every The Borg model can do a few important things:

train(): This trains the model on a training set.

call: Run the emulator for a set of data/labels, e.g., out = model(input).

save(): Save the model to a file.

load(): Load a model from a file.

Training a Model¶

We use a training set with data and labels to train the models. The dimensions of the training_labels and training_data should be:

training_labels: [# training data, # labels]

training_data: [# training data, # features]

You can specify various parameters of the model. For example, num_neurons, the number of neurons in the hidden layer(s). You can also specify some training parameters, such as how quickly the model learns (learning rate), the chunks of training data that are fed through the model simultaneously during the training process (batch size), and the number of iterations to train (num_steps).

It is substantially faster to train an ANN using a GPU with the cuda=True setting. By default, cuda=True so a model can be trained on a regular CPU.

Classifier¶

A classifier takes data with various features and returns a probability of it being a member of various classes.

from theborg.classifier import Classifier
cmodel = Classifier()
cmodel.train(training_labels=labels,training_data=data)
probs = cmodel(data)

The input data when calling the model will have dimensions of [# of features] and the output will have probabilities [# of labels] that sum to one.

Now you can save the model to a file.

cmodel.save('my_classifier.pkl')

There are two classifiers. A simple classifier Classifier using a serial model, and a classifier using convolutional layers CNNClassifier.

Emulator¶

An emulator simulates real data using only a few labels. For example, a stellar spectrum is simulated using just temperature, surface gravity and metallicity. Emulator can be used to emulate a wide variety of data.

from theborg.emulator import Emulator
emodel = Emulator()
emodel.train(training_labels=labels,training_data=data)
spec = emodel(labels)

The input data when calling the model will have dimensions of [# of labels] and the output will have dimensions of [# of features].

Navigation