train: Gradient Descent trainer

In IshmaelBelghazi/gpuClassifieR: gpuClassifieR C/CUDA linear classifier for R

Description

Gradient Descent trainer

Usage

train(object, feats, targets, decay = NULL, step_size = NULL,
  max_iter = NULL, verbose = FALSE, tol = 1e-06, backend = "R", ...)

## S3 method for class 'model.spec'
train(object, feats, targets, decay = NULL,
  step_size = NULL, max_iter = NULL, verbose = FALSE, tol = 1e-06,
  backend = "R", ...)

Arguments

object	Linear classifier specification object.
feats	Numeric Matrix of features. Follows the usual convention of having one example per row. For a model with M features and dataset with N examples the matrix should be N \times K
targets	Numeric matrix of one hot encoded target. Follows the usual convention of having one target per row. For a model with K classes and a dataset with N examples the matrix should be N \times K
decay	Numeric scalar. Tikhonov regularization coefficient (weight decay). Should be a non-negative real number.
step_size	Numeric scalar. Initial, gradient descent step size. If the current cost is worst that the previous, model parameters will not be update and the step_size will be divided by 2. If the current cost not the worse than the previous, parameters will be updated and the step_size will be multiplied by 1.1 .
max_iter	Numeric Scalar. Maximum number of iterations allowed. Inf will keep the training going until the number of the current gradient Frobenius norm is less than tol.
verbose	Logical scalar. If TRUE, iteration number, current cost and step_size will be printed to the standard output.
tol	Numeric scalar. Assumes convergence when the gradient Frobenius norm is less than tol.
backend	Computation back-end ('R', 'C', or 'CUDA')
...	other arguments passed to specific methods

Value

model.spec object

Methods (by class)

• model.spec: Gradient Descent trainer for linear classifier specification model objects.

Author(s)

Mohamed Ishmael Diwan Belghazi

Examples

# Train model on single example of MIST and a single iteration
# Generate random initial weights
w_init <- matrix(rnorm(784 * 10), 784, 10)
# construct model
linear_classifier <- Classifier(weights_init=w_init)
# Fetch training variables
feats <- mini_mnist$train$images[1, , drop=FALSE]
targets <- mini_mnist$train$labels[1, , drop=FALSE]
# Specifying training parameters
step_size <- 0.01
decay <- 0.0001
max_iter <- 1
tol <- 1e-6
verbose <- FALSE
# Train model one a single example using the three back-ends
linear_classifier_R <- train(linear_classifier, feats, targets, decay,
step_size, max_iter, verbose, tol, backend='R')
linear_classifier_C <- train(linear_classifier, feats, targets, decay,
step_size, max_iter, verbose, tol, backend='C')
linear_classifier_CUDA <- train(linear_classifier, feats, targets, decay,
step_size, max_iter, verbose, tol, backend='CUDA')

IshmaelBelghazi/gpuClassifieR documentation built on May 7, 2019, 6:45 a.m.