cuda.ml (version 0.3.2)

cuda_ml_sgd: Train a MBSGD linear model.

Description

Train a linear model using mini-batch stochastic gradient descent.

Usage

cuda_ml_sgd(x, ...)
# S3 method for default
cuda_ml_sgd(x, ...)
# S3 method for data.frame
cuda_ml_sgd(
  x,
  y,
  fit_intercept = TRUE,
  loss = c("squared_loss", "log", "hinge"),
  penalty = c("none", "l1", "l2", "elasticnet"),
  alpha = 1e-04,
  l1_ratio = 0.5,
  epochs = 1000L,
  tol = 0.001,
  shuffle = TRUE,
  learning_rate = c("constant", "invscaling", "adaptive"),
  eta0 = 0.001,
  power_t = 0.5,
  batch_size = 32L,
  n_iters_no_change = 5L,
  ...
)
# S3 method for matrix
cuda_ml_sgd(
  x,
  y,
  fit_intercept = TRUE,
  loss = c("squared_loss", "log", "hinge"),
  penalty = c("none", "l1", "l2", "elasticnet"),
  alpha = 1e-04,
  l1_ratio = 0.5,
  epochs = 1000L,
  tol = 0.001,
  shuffle = TRUE,
  learning_rate = c("constant", "invscaling", "adaptive"),
  eta0 = 0.001,
  power_t = 0.5,
  batch_size = 32L,
  n_iters_no_change = 5L,
  ...
)
# S3 method for formula
cuda_ml_sgd(
  formula,
  data,
  fit_intercept = TRUE,
  loss = c("squared_loss", "log", "hinge"),
  penalty = c("none", "l1", "l2", "elasticnet"),
  alpha = 1e-04,
  l1_ratio = 0.5,
  epochs = 1000L,
  tol = 0.001,
  shuffle = TRUE,
  learning_rate = c("constant", "invscaling", "adaptive"),
  eta0 = 0.001,
  power_t = 0.5,
  batch_size = 32L,
  n_iters_no_change = 5L,
  ...
)
# S3 method for recipe
cuda_ml_sgd(
  x,
  data,
  fit_intercept = TRUE,
  loss = c("squared_loss", "log", "hinge"),
  penalty = c("none", "l1", "l2", "elasticnet"),
  alpha = 1e-04,
  l1_ratio = 0.5,
  epochs = 1000L,
  tol = 0.001,
  shuffle = TRUE,
  learning_rate = c("constant", "invscaling", "adaptive"),
  eta0 = 0.001,
  power_t = 0.5,
  batch_size = 32L,
  n_iters_no_change = 5L,
  ...
)

Arguments

x

Depending on the context:

* A __data frame__ of predictors. * A __matrix__ of predictors. * A __recipe__ specifying a set of preprocessing steps * created from [recipes::recipe()]. * A __formula__ specifying the predictors and the outcome.

...

Optional arguments; currently unused.

y

A numeric vector (for regression) or factor (for classification) of desired responses.

fit_intercept

If TRUE, then the model tries to correct for the global mean of the response variable. If FALSE, then the model expects data to be centered. Default: TRUE.

loss

Loss function, must be one of "squared_loss", "log", "hinge".

penalty

Type of regularization to perform, must be one of "none", "l1", "l2", "elasticnet".

- "none": no regularization. - "l1": perform regularization based on the L1-norm (LASSO) which tries to minimize the sum of the absolute values of the coefficients. - "l2": perform regularization based on the L2 norm (Ridge) which tries to minimize the sum of the square of the coefficients. - "elasticnet": perform the Elastic Net regularization which is based on the weighted averable of L1 and L2 norms. Default: "none".

alpha

Multiplier of the penalty term. Default: 1e-4.

l1_ratio

The ElasticNet mixing parameter, with 0 <= l1_ratio <= 1. For l1_ratio = 0 the penalty is an L2 penalty. For l1_ratio = 1 it is an L1 penalty. For 0 < l1_ratio < 1, the penalty is a combination of L1 and L2. The penalty term is computed using the following formula: penalty = alpha * l1_ratio * ||w||_1 + 0.5 * alpha * (1 - l1_ratio) * ||w||^2_2 where ||w||_1 is the L1 norm of the coefficients, and ||w||_2 is the L2 norm of the coefficients.

epochs

The number of times the model should iterate through the entire dataset during training. Default: 1000L.

tol

Threshold for stopping training. Training will stop if (loss in current epoch) > (loss in previous epoch) - tol. Default: 1e-3.

shuffle

Whether to shuffles the training data after each epoch. Default: True.

learning_rate

Must be one of "constant", "invscaling", "adaptive".

- "constant": the learning rate will be kept constant. - "invscaling": (learning rate) = (initial learning rate) / pow(t, power_t) where t is the number of epochs and power_t is a tunable parameter of this model. - "adaptive": (learning rate) = (initial learning rate) as long as the training loss keeps decreasing. Each time the last n_iter_no_change consecutive epochs fail to decrease the training loss by tol, the current learning rate is divided by 5. Default: "constant".

eta0

The initial learning rate. Default: 1e-3.

power_t

The exponent used in the invscaling learning rate calculations.

batch_size

The number of samples that will be included in each batch. Default: 32L.

n_iters_no_change

The maximum number of epochs to train if there is no imporvement in the model. Default: 5.

formula

A formula specifying the outcome terms on the left-hand side, and the predictor terms on the right-hand side.

data

When a __recipe__ or __formula__ is used, data is specified as a __data frame__ containing the predictors and (if applicable) the outcome.

Value

A linear model that can be used with the 'predict' S3 generic to make predictions on new data points.

Examples

# NOT RUN {
library(cuda.ml)

model <- cuda_ml_sgd(
  mpg ~ ., mtcars,
  batch_size = 4L, epochs = 50000L,
  learning_rate = "adaptive", eta0 = 1e-5,
  penalty = "l2", alpha = 1e-5, tol = 1e-6,
  n_iters_no_change = 10L
)

preds <- predict(model, mtcars[names(mtcars) != "mpg"])
print(all.equal(preds$.pred, mtcars$mpg, tolerance = 0.09))
# }