SubNNP

class hdnnpy.model.models.SubNNP(element, n_feature, hidden_layers, n_property)

Bases: chainer.link.Chain

Feed-forward neural network representing one element or atom.

element is registered as a persistent value.

It consists of repetition of fully connected layer and activation function.

Weight initializer is chainer.initializers.HeNormal.

Parameters:

• element (str) – Element symbol represented by an instance.
• n_feature (int) – Number of nodes of input layer.
• hidden_layers (list [tuple [int, str]]) – A neural network structure. Last one is output layer, and the remains are hidden layers. Each element is a tuple (# of nodes, activation function), for example (50, 'sigmoid'). Only activation functions implemented in chainer.functions can be used.
• n_property (int) – Number of nodes of output layer.

__len__()

Return the number of hidden_layers.

differentiate(x, enable_double_backprop)

Calculate derivative of the output data w.r.t. input data.

Parameters:

• x (Variable) – Input data which has the shape (n_sample, n_input).
• enable_double_backprop (bool) – Passed to chainer.grad() to determine whether to create more deep calculation graph or not.

feedforward(x)

Propagate input data in a feed-forward way.

Parameters:x (Variable) – Input data which has the shape (n_sample, n_input).

second_differentiate(x, enable_double_backprop)

Calculate 2nd derivative of the output data w.r.t. input data.

Parameters:

• x (Variable) – Input data which has the shape (n_sample, n_input).
• enable_double_backprop (bool) – Passed to chainer.grad() to determine whether to create more deep calculation graph or not.