Neural networks

class untwist.neuralnetworks.Layer(input, n_in, n_out, W=None, b=None, activation=<theano.tensor.elemwise.Elemwise object>)

Generic feed-forward layer (can be hidden or output)

class untwist.neuralnetworks.Activations

Some activation functions not available in theano.

classmethod ReLU()

classmethod ReLU2()

class untwist.neuralnetworks.MLP(input_size, output_size, hidden_sizes, activation=<theano.tensor.elemwise.Elemwise object>)

Multi-Layer Perceptron network, specified by the size of the input, the output and an arbitrary number of hidden layers.

load(fname)

save(fname)

class untwist.neuralnetworks.SGD(mlp, learning_rate=0.1, momentum=0.5, batch_size=100, iterations=100, patience=0, rate_decay_th=0.1)

Stochastic gradient descent algorithm for training MLP instances Includes momentum, learning rate scheduling, early stopping.

predict(data)

Get mlp prediction from observation.

train(dataset)

Train the MLP passed to the constructor using a Dataset.