NeuralNetwork

class safe_learning.NeuralNetwork(layers, nonlinearities, output_scale=1.0, use_bias=True, name='neural_network')

A simple neural network.

The neural network also exposes its Lipschitz constant as NeuralNetwork.lipschitz.

Parameters:

layers : list

A list of layer sizes, [l1, l2, l3, .., ln]. l1 corresponds to the input dimension of the neural network, while ln is the output dimension.

nonlinearities : list

A list of nonlinearities applied after each layer. Can be None if no nonlinearity should be applied.

output_scale : float, optional

A constant scaling factor applied to the neural network output.

use_bias : bool, optional

A boolean determining whether bias terms are included in each hidden layer; bias terms are never used in the output layer.

name : string, optional

Attributes:

parameters

Return the variables within the current scope.

scope_name

Methods

__call__(self, \*args, \*\*kwargs)	Evaluate the function using the template to ensure variable sharing.
build_evaluation(self, points)	Build the evaluation graph.
copy_parameters(self, other_instance)	Copy over the parameters of another instance.
lipschitz(self)	Return the Lipschitz constant as a Tensor.

build_evaluation(self, points)

Build the evaluation graph.

copy_parameters(self, other_instance)

Copy over the parameters of another instance.

lipschitz(self)

Return the Lipschitz constant as a Tensor.

This assumes that only contractive nonlinearities are used! Examples are ReLUs and Sigmoids.

Returns:

lipschitz : Tensor

The Lipschitz constant of the neural network.

parameters

Return the variables within the current scope.