get_grad: Computes model gradient

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

Description

Computes model gradient

Usage

get_grad(object, feats, targets, decay = NULL, backend = "R", ...)

## S3 method for class 'model.spec'
get_grad(object, feats, targets, decay = NULL,
  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.
backend	Computation back-end ('R', 'C', or 'CUDA')
...	other arguments passed to specific methods

Value

Numeric Matrix of gradients. One for each class. Gradient are arrayed in columns. For a model with M features and K classes the matrix should be M \times K

Methods (by class)

• model.spec: Computes model gradient for linear classifier specification objects

Author(s)

Mohamed Ishmael Diwan Belghazi

Examples

# 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
targets <- mini_mnist$train$labels
# Set decay coefficient
decay <- 0.01
# compute gradient at the training set using the three back-ends
gradient_R <- get_grad(linear_classifier, feats, targets, decay, 'R')
gradient_C <- get_grad(linear_classifier, feats, targets, decay, 'C')
gradient_CUDA <- get_grad(linear_classifier, feats, targets, decay, 'CUDA')

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