Skip to contents

Generate architecture list for Deep Neural Network and Convolutional Neural Network

Source: R/generate_arch_list.R
generate_arch_list.Rd

This function generates a list of architectures for either a Deep Neural Netwokd (DNN) or a Convolutional Neural Network (CNN).

Usage

generate_arch_list(
  type,
  number_of_features,
  number_of_outputs,
  n_layers = c(1, 2),
  n_neurons = c(7),
  sample_size = c(11, 11),
  number_of_fc_layers = 1,
  fc_layers_size = c(14),
  conv_layers_kernel = 3,
  conv_layers_stride = 1,
  conv_layers_padding = 0,
  pooling = NULL,
  batch_norm = TRUE,
  dropout = 0
)

Arguments

type

string. Specifies the type the network. The valid inputs "dnn" and "cnn".

number_of_features

numeric. Value that specifies the number of features in the dataset.

number_of_outputs

numeric. Value that specifies the number of outputs.

n_layers

numeric. Vector that specifies the number of layers in the networks. Default value are 1 and 2.

n_neurons

vector. Specifies the number of neurons each layer. Default 7.

sample_size

vector. Specifies the size. Default c(11, 11)

number_of_fc_layers

numeric. Specifies the number of fully connected layers. Default 1.

fc_layers_size

vector. Specifies the size of the fully connected layers. Default 14.

conv_layers_kernel

numeric. Specifies the kernel size for layers. Default 3.

conv_layers_stride

numeric. Specifies the stride for the convolutional layers. Default 1.

conv_layers_padding

numeric. Specifies the padding for the convolutional layers. Default 0.

pooling

numeric. Specifies 2D average pooling kernel size. Default NULL

batch_norm

logical. Specifies whether batch normalization is included in the architecture. Default TRUE.

dropout

Numeric. The probability (p) of randomly zeroing elements of the input tensor during training to prevent overfitting. Must be between 0 (no dropout) and 1 (all inputs zeroed). Default is 0 (no dropout).

Value

A list containing:

  • arch_list: a list of generated architectures.

  • arch_dict: a list of architecture dictionaries.

See also

generate_dnn_architecture, generate_cnn_architecture

Examples

if (FALSE) {
# Generating architectures for DNN, using batch normalization and dropout
dnn_archs <- generate_arch_list(
  type = "dnn",
  number_of_features = 4,
  number_of_outputs = 1,
  n_layers = c(2, 3, 4),
  n_neurons = c(8, 16, 32, 64),
  batch_norm = TRUE,
  dropout = 0.2
)

dnn_archs$arch_dict # Matrices describing the networks
length(dnn_archs$arch_list) # Generated 336 DNN architectures

# Generating architectures for CNN, using batch normalization, dropout and average pooling
# Note that arguments meaning change with the context

cnn_archs <- generate_arch_list(
  type = "cnn",
  number_of_features = 4,
  number_of_outputs = 1,
  n_layers = c(2, 3, 4), # now convolutional layers
  n_neurons = c(8, 16, 32, 64),
  sample_size = c(11, 11),
  number_of_fc_layers = c(2, 4), # fully connected layers
  fc_layers_size = c(16, 8),
  conv_layers_kernel = 3,
  conv_layers_stride = 1,
  conv_layers_padding = 0,
  pooling = 1,
  batch_norm = TRUE,
  dropout = 0.2
)

cnn_archs$arch_dict # Matrices describing the networks
length(cnn_archs$arch_list) # Generated 6720 CNN architectures

# The list size can be easily and greatly reduced with select_arch_list

dnn_archs_redux <- dnn_archs %>% select_arch_list(
  type = c("dnn"),
  method = "percentile",
  n_samples = 1,
  min_max = TRUE # Keep the network with the minimum and maximum number of parameters
)

length(dnn_archs_redux$arch_list) # from 336 to 29 architectures

cnn_archs_redux <- cnn_archs %>% select_arch_list(
  type = c("cnn"),
  method = "percentile",
  n_samples = 1,
  min_max = TRUE # Keep the network with the minimum and maximum number of parameters
)

length(cnn_archs_redux$arch_list) # from 6720 to 77 architectures
}