Select architectures for Convolutional Neural Network or Deep Neural Network

Usage

select_arch_list(
  arch_list,
  type = c("dnn", "cnn"),
  method = "percentile",
  n_samples = 1,
  min_max = TRUE
)

Arguments

arch_list: list. Containing Convolutional Neural Network or Deep Neural Network architectures.
type: character. Indicating the type of network. Options are "dnn" or "cnn".
method: character. Indicating the method to select architectures. Default is "percentile".
n_samples: integer. Specifying the number of samples to select per group. Default is 1.
min_max: logical. If TRUE, include networks with minimal and maximal parameters.

Value

A list with:

arch_list: a list containing torch neural networks
arch_dict: a list of matrices describing the structure of those networks
changes: a tibble with information about neural networks name changes, number of parameters and deepness

Examples

if (FALSE) {
# Generate some big list of architectures combining all argument values
big_arch_list <- generate_arch_list(
  type = "dnn",
  number_of_features = 4,
  number_of_outputs = 1,
  n_layers = seq(from = 2, to = 6, by = 1),
  n_neurons = c(8, 16, 32, 64)
)

length(big_arch_list$arch_list) # 5456 architectures!

# It can be reduced sampling network architectures by its parameters number

reduced_arch_list <- big_arch_list %>% select_arch_list(
  type = c("dnn"),
  method = "percentile",
  n_samples = 1, # Keep at least one of each deepness
  min_max = TRUE # Keep the network with the minimum and maximum number of parameters
)

length(reduced_arch_list$arch_list) # from 5456 to 92 architectures!!

# See architectures names, deepness and number of parameters
reduced_arch_list$changes
}