Consequent Functions

This module defines the consequent functions used in the toolbox. These functions define the output of each rule in neuro-fuzzy models.

All classes below inherit from the base class ConsequentFunction, which defines the common interface and serves as a reference for future implementations.

Note

Currently, only the standard linear consequent function of the ANFIS model is implemented. Adding new consequent function types would require non-trivial changes to the model and training algorithm implementations.

class neuro_fuzzy_toolbox.func.consequent.Linear_CF(*args: Any, **kwargs: Any)[source]

Bases: ConsequentFunction

Linear consequent function.

Computes the output of the neuro-fuzzy network from the input features and rule consequents as a linear combination, defined as:

\[O_j = \sum_{i=1}^{n} (c_{i,j} \cdot x_i) + c_{n+1,j}\]
where:

  • \(O_j\) is the \(j\)-th output of the consequent layer of an ANFIS model (associated to the \(j\)-th rule).

  • \(x_i\) is the \(i\)-th feature of an input sample \(x\) of size \(n\).

  • \(c_{i,j}\) is the \(i\)-th consequent parameter associated with the \(j\)-th rule of an ANFIS model.

forward(x, consequents, weights)[source]

Forward pass of the linear consequent function.

Parameters:

  • x (torch.Tensor) – Input tensor of shape (batch_size, input_size) containing the input features.

  • consequents (torch.Tensor) – Tensor of shape (outputs, rules, input_size + 1) containing the consequent parameters, where rules is the number of fuzzy rules and outputs is the number of model outputs.

  • weights (torch.Tensor) – Tensor of shape (batch_size, rules) containing the normalized firing levels for each rule.

Returns:

Tensor of shape (outputs, batch_size, rules) containing the weighted rule outputs.

Return type:

torch.Tensor

get_consequents_outputs(x, consequents)[source]

Returns the individual rule outputs without weighting by normalized firing levels.

Parameters:

  • x (torch.Tensor) – Input tensor of shape (batch_size, input_size) containing the input features.

  • consequents (torch.Tensor) – Tensor of shape (outputs, rules, input_size + 1) containing the consequent parameters, where rules is the number of fuzzy rules and outputs is the number of model outputs.

Returns:

Tensor of shape (outputs, batch_size, rules) containing the unweighted output of each rule, without multiplication by the normalized firing levels.

Return type:

torch.Tensor

random_consequents(outputs, rules, input_size, dtype)[source]

Initializes the consequent parameters randomly in the range [-1, 1].

Parameters:

  • outputs (int) – Number of model outputs.

  • rules (int) – Number of fuzzy rules.

  • input_size (int) – Number of input features.

  • dtype (torch.dtype) – Data type for the returned tensor.

Returns:

Tensor of shape (outputs, rules, input_size + 1) containing randomly initialized consequent parameters.

Return type:

torch.Tensor