Internal Neural Network Interfaces#

class drdmannturb.TauNet(n_layers: int = 2, hidden_layer_size: int = 3, learn_nu: bool = True)[source]#

Classical implementation of neural network that learns the eddy lifetime function \(\tau(\boldsymbol{k})\). A SimpleNN and Rational network comprise this class. The network widths are determined by a single integer and thereafter the networks have hidden layers of only that width.

The objective is to learn the function

\[\tau(\boldsymbol{k})=\frac{T|\boldsymbol{a}|^{\nu-\frac{2}{3}}}{\left(1+|\boldsymbol{a}|^2\right)^{\nu / 2}}, \quad \boldsymbol{a}=\boldsymbol{a}(\boldsymbol{k}),\]

where

\[\boldsymbol{a}(\boldsymbol{k}) = \operatorname{abs}(\boldsymbol{k}) + \mathrm{NN}(\operatorname{abs}(\boldsymbol{k})).\]

This class implements the simplest architectures which solve this problem.

Parameters:

n_layers (int, optional) – Number of hidden layers, by default 2
hidden_layer_size (int, optional) – Size of the hidden layers, by default 3
learn_nu (bool, optional) – If true, learns also the exponent \(\nu\), by default True

forward(k: Tensor) → Tensor[source]#

Forward method implementation. Evaluates

Parameters:: k (torch.Tensor) – Wave vector input
Returns:: torch.Tensor – Output of forward pass of neural network.
Return type:: Tensor

class drdmannturb.CustomNet(n_layers: int = 2, hidden_layer_sizes: int | list[int] = [10, 10], activations: List[Module] = [ReLU(), ReLU()], learn_nu: bool = True)[source]#

A more versatile version of the tauNet. The objective is the same: to learn the eddy lifetime function \(\tau(\boldsymbol{k})\) in the same way. This class allows for neural networks of variable widths and different kinds of activation functions used between layers.

Parameters:

n_layers (int, optional) – Number of hidden layers, by default 2
hidden_layer_sizes (Union[int, list[int]]) – Sizes of each layer; by default [10, 10].
activations (List[nn.Module], optional) – List of activation functions to use, by default [nn.ReLU(), nn.ReLU()]
learn_nu (bool, optional) – If true, learns also the exponent \(\nu\), by default True

forward(k: Tensor) → Tensor[source]#

Forward method implementation

Parameters:: k (torch.Tensor) – Input wavevector domain.
Returns:: torch.Tensor – Output of forward pass of neural network
Return type:: Tensor

class drdmannturb.SimpleNN(nlayers: int = 2, inlayer: int = 3, hlayer: int = 3, outlayer: int = 3)[source]#

A simple feed-forward neural network consisting of n layers with a ReLU activation function. The default initialization is to random noise of magnitude 1e-9.

Parameters:

nlayers (int, optional) – Number of layers to use, by default 2
inlayer (int, optional) – Number of input features, by default 3
hlayer (int, optional) – Number of hidden layers, by default 3
outlayer (int, optional) – Number of output features, by default 3

forward(x: Tensor) → Tensor[source]#

Forward method implementation

Parameters:: x (torch.Tensor) – Network input
Returns:: torch.Tensor – Network output
Return type:: Tensor

class drdmannturb.Rational(learn_nu: bool = True)[source]#

Learnable rational kernel; a neural network that learns the rational function

\[\tau(\boldsymbol{k})=\frac{T|\boldsymbol{a}|^{\nu-\frac{2}{3}}}{\left(1+|\boldsymbol{a}|^2\right)^{\nu / 2}}, \quad \boldsymbol{a}=\boldsymbol{a}(\boldsymbol{k}),\]

specifically, the neural network part of the augmented wavevector

\[\mathrm{NN}(\operatorname{abs}(\boldsymbol{k})).\]

Parameters:: learn_nu (bool, optional) – Indicates whether or not the exponent nu should be learned also; by default True

forward(x: Tensor) → Tensor[source]#

Forward method implementation

Parameters:: x (torch.Tensor) – Network input
Returns:: torch.Tensor – Network output
Return type:: Tensor

class drdmannturb.CustomMLP(hlayers: list[int], activations: list[Module], inlayer: int = 3, outlayer: int = 3)[source]#

Feed-forward neural network with variable widths of layers and activation functions. Useful for DNN configurations and experimentation with different activation functions.

Parameters:

hlayers (list) – list specifying widths of hidden layers in NN
activations (list[nn.Module]) – list specifying activation functions for each hidden layer
inlayer (int, optional) – Number of input features, by default 3
outlayer (int, optional) – Number of features to output, by default 3

forward(x: Tensor) → Tensor[source]#

Forward method implementation

Parameters:: x (torch.Tensor) – Input to the network
Returns:: torch.Tensor – Output of the network
Return type:: Tensor