Package 'rkaf'

Extract KAF Fourier parameters

Description

Extracts the trainable Random Fourier Feature weights and biases from a KAF model.

Usage

extract_fourier_params(object, layer = NULL)

Arguments

object	A fitted object returned by kaf_fit(), or a raw nn_kaf torch module.
layer	Optional integer. If supplied, only extract parameters from this layer.

Value

A data frame with Fourier weights and biases.

---

Extract KAF branch scales

Description

Extracts the learned GELU/base and Fourier branch scales from each KAF layer.

Usage

extract_kaf_scales(object)

Arguments

object

A fitted object returned by kaf_fit(), or a raw nn_kaf torch module.

Value

A data frame with one row per layer-feature pair.

---

Create a Kolmogorov-Arnold Fourier Network

Description

Convenience wrapper around nn_kaf() using a more user-facing API.

Usage

kaf(
  input_dim,
  output_dim = 1,
  hidden = c(64, 64),
  num_grids = 16,
  dropout = 0,
  use_layernorm = TRUE,
  activation_expectation = 1.64,
  fourier_init_scale = 0.01
)

Arguments

input_dim	Integer. Number of input features.
output_dim	Integer. Number of output dimensions.
hidden	Integer vector. Hidden layer sizes.
num_grids	Integer. Number of Fourier frequencies per KAF layer.
dropout	Numeric. Dropout probability.
use_layernorm	Logical. Whether to apply layer normalization.
activation_expectation	Numeric. Scaling constant for Fourier initialization.
fourier_init_scale	Numeric. Initial scale of the Fourier branch.

Value

A torch KAF network.

---

Fit a Kolmogorov-Arnold Fourier Network

Description

Fits a KAF model for regression using mean squared error loss.

Usage

kaf_fit(
  x,
  y,
  task = c("auto", "regression", "binary", "multiclass"),
  hidden = c(64, 64),
  num_grids = 16,
  dropout = 0,
  use_layernorm = TRUE,
  fourier_init_scale = 0.01,
  epochs = 1000,
  lr = 0.001,
  batch_size = NULL,
  shuffle = TRUE,
  validation_split = 0,
  x_val = NULL,
  y_val = NULL,
  weight_decay = 0,
  standardize_x = TRUE,
  standardize_y = NULL,
  patience = NULL,
  verbose = TRUE,
  print_every = 100,
  seed = NULL,
  restore_best = TRUE,
  min_delta = 0
)

Arguments

x	Matrix, data frame, vector, or 2D torch tensor of predictors.
y	Vector, matrix, data frame, or torch tensor of targets.
task	Character. One of "auto", "regression", "binary", or "multiclass". With "auto", factor, character, and logical targets are treated as classification; numeric targets are treated as regression.
hidden	Integer vector. Hidden layer sizes.
num_grids	Integer. Number of Fourier frequencies per KAF layer.
dropout	Numeric. Dropout probability.
use_layernorm	Logical. Whether to apply layer normalization.
fourier_init_scale	Numeric. Initial scale of the Fourier branch.
epochs	Integer. Maximum number of training epochs.
lr	Numeric. Learning rate.
batch_size	Optional integer. Mini-batch size. If NULL, full-batch training is used.
shuffle	Logical. Whether to shuffle training rows each epoch.
validation_split	Numeric in ⁠[0, 1)⁠. Fraction of rows to reserve for validation. Ignored if x_val and y_val are supplied.
x_val	Optional validation predictors.
y_val	Optional validation targets.
weight_decay	Numeric. Adam weight decay.
standardize_x	Logical. Whether to standardize predictors using the training-set mean and standard deviation.
standardize_y	Logical or NULL. Whether to standardize regression targets using the training-set mean and standard deviation. If NULL, targets are standardized for regression and not standardized for classification. Predictions are automatically transformed back to the original target scale.
patience	Optional integer. Number of epochs without improvement before early stopping.
verbose	Logical. Whether to print progress.
print_every	Integer. Print frequency.
seed	Optional integer random seed.
restore_best	Logical. Whether to restore the best observed model state after training.
min_delta	Numeric. Minimum loss improvement required to update the best model state.

Value

An object of class "kaf_fit".

---

Fit a KAF model using an R formula

Description

Fits a Kolmogorov-Arnold Fourier Network using a formula and data frame. This is a convenience wrapper around kaf_fit() for regression tasks.

Usage

kaf_fit_formula(
  formula,
  data,
  include_intercept = FALSE,
  na.action = stats::na.omit,
  ...
)

Arguments

formula	A model formula, such as y ~ x1 + x2.
data	A data frame.
include_intercept	Logical. Whether to keep the model-matrix intercept column. Defaults to FALSE.
na.action	Missing-value handling function passed to model.frame().
...	Additional arguments passed to kaf_fit().

Value

An object of class "kaf_fit".

---

Kolmogorov-Arnold Fourier Network

Description

Torch module implementing a stacked Kolmogorov-Arnold Fourier Network.

Usage

nn_kaf(
  layers,
  num_grids = 8,
  dropout = 0,
  use_layernorm = TRUE,
  activation_expectation = 1.64,
  fourier_init_scale = 0.01
)

Arguments

layers	Integer vector. Network architecture, including input and output dimensions. For example, c(10, 64, 64, 1).
num_grids	Integer. Number of Fourier frequencies per KAF layer.
dropout	Numeric. Dropout probability.
use_layernorm	Logical. Whether to apply layer normalization before the Fourier feature block.
activation_expectation	Numeric. Scaling constant for Fourier initialization.
fourier_init_scale	Numeric. Initial scale of the Fourier component.

Value

A torch nn_module.

---

Kolmogorov-Arnold Fourier Layer

Description

Torch module implementing one KAF layer using a hybrid GELU and trainable Random Fourier Features activation.

Usage

nn_kaf_layer(
  input_dim,
  output_dim,
  num_grids = 8,
  dropout = 0,
  use_layernorm = TRUE,
  activation_expectation = 1.64,
  fourier_init_scale = 0.01
)

Arguments

input_dim	Integer. Input dimension.
output_dim	Integer. Output dimension.
num_grids	Integer. Number of Fourier frequencies.
dropout	Numeric. Dropout probability.
use_layernorm	Logical. Whether to apply layer normalization before the Fourier feature block.
activation_expectation	Numeric. Scaling constant for Fourier initialization.
fourier_init_scale	Numeric. Initial scale of the Fourier component.

Details

The layer computes a feature-wise hybrid activation of the form:

$z = \alpha \odot GELU(x) + \beta \odot RFF(x)$

followed by a linear projection to the output dimension.

Value

A torch nn_module.

---

Random Fourier Features Layer

Description

Torch module implementing trainable random Fourier features for Kolmogorov-Arnold Fourier Networks.

Usage

nn_random_fourier_features(
  input_dim,
  num_grids = 8,
  dropout = 0,
  activation_expectation = 1.64
)

Arguments

input_dim	Integer. Input dimension.
num_grids	Integer. Number of Fourier frequencies.
dropout	Numeric. Dropout probability.
activation_expectation	Numeric. Scaling constant for initialization.

Value

A torch nn_module.

---

Plot KAF branch scales

Description

Visualizes the learned base/GELU and Fourier branch scales for a selected KAF layer.

Usage

plot_kaf_scales(object, layer = 1, type = c("ratio", "branch"), ...)

Arguments

object	A fitted object returned by kaf_fit(), or a raw nn_kaf torch module.
layer	Integer. Layer to inspect.
type	Character. Either "ratio" or "branch".
...	Additional arguments passed to base plotting functions.

Value

Invisibly returns the extracted scale data.

---

Plot a fitted KAF model

Description

Plot a fitted KAF model

Usage

## S3 method for class 'kaf_fit'
plot(x, type = c("loss", "fit"), newdata = NULL, y = NULL, ...)

Arguments

x	A fitted object returned by kaf_fit().
type	Character. Either "loss" or "fit".
newdata	Optional predictors used when type = "fit".
y	Optional observed target values used when type = "fit".
...	Additional arguments passed to base plotting functions.

Value

Invisibly returns x.

---

Predict from a fitted KAF model

Description

Predict from a fitted KAF model

Usage

## S3 method for class 'kaf_fit'
predict(
  object,
  newdata,
  type = c("response", "prob", "class", "link"),
  threshold = 0.5,
  as_tensor = FALSE,
  ...
)

Arguments

object	A fitted object returned by kaf_fit().
newdata	Matrix, data frame, vector, or 2D torch tensor.
type	Character. Prediction type. "response" returns regression predictions for regression, probabilities for classification. "prob" returns probabilities for classification. "class" returns predicted classes. "link" returns raw model logits/outputs.
threshold	Numeric. Classification threshold used for binary class predictions.
as_tensor	Logical. If TRUE, return a torch tensor where supported.
...	Unused.

Value

Predictions as a vector, matrix, factor, or torch tensor.

---

Print a fitted KAF model

Description

Print a fitted KAF model

Usage

## S3 method for class 'kaf_fit'
print(x, ...)

Arguments

x	A fitted object returned by kaf_fit().
...	Unused.

Value

Invisibly returns x.

---