Skip to contents

Fit and validate Deep Neural Network model

Source: R/fit_abund_dnn.R
fit_abund_dnn.Rd

Fit and validate Deep Neural Network model

Usage

fit_abund_dnn(
  data,
  response,
  predictors,
  predictors_f = NULL,
  partition,
  hold_out_set = NULL,
  predict_part = FALSE,
  learning_rate = 0.01,
  weight_decay = 0,
  n_epochs = 10,
  batch_size = 32,
  validation_patience = 2,
  fitting_patience = 5,
  optimizer = torch::optim_adamw,
  loss_function = torch::nn_l1_loss,
  custom_architecture = NULL,
  verbose = TRUE,
  learning_monitor = FALSE
)

Arguments

data

tibble or data.frame. Database with response, predictors, and partition values

response

character. Column name with species abundance.

predictors

character. Vector with the column names of quantitative predictor variables (i.e. continuous variables). Usage predictors = c("temp", "precipt", "sand")

predictors_f

character. Vector with the column names of qualitative predictor variables (i.e. ordinal or nominal variables type). Usage predictors_f = c("landform")

partition

character. Column name with training and validation partition groups.

predict_part

logical. Save predicted abundance for testing data. Default = FALSE

learning_rate

numeric. The size of the step taken during the optimization process. Default = 0.01

weight_decay

numeric. The regularization strength: 0 means no penalty, while higher values (e.g. 0.01) apply stronger shrinkage to the weights during training. Default is 0

n_epochs

numeric. Max number of times the learning algorithm will work through the training set. Default = 10

batch_size

numeric. A batch is a subset of the training set used in a single iteration of the training process. The size of each batch is referred to as the batch size. Default = 32

validation_patience

numerical. An integer indicating the number of epochs without loss improvement tolerated by the algorithm in the validation process. If the patience limit is exceeded, the training ends. Default 2

fitting_patience

numerical. The same as validation_patience, but in the final model fitting process. Default 5

optimizer

a torch_optimizer_generator. The optimizer to be used in model fitting. Default is torch::optim_adamw.

loss_function

a torch nn_loss. The loss function to be used in model fitting. Default is torch::nn_l1_loss.

custom_architecture

a Torch nn_module_generator object or a generate_dnn_architecture output. A neural network architecture to be used instead of the internal default one. Default NULL

verbose

logical. If FALSE, disables all console messages. Default TRUE

learning_monitor

logical. If TRUE, the function will return a tibble containing loss values over training epochs. It is useful to create learning/convergence plots.

Value

A list object with:

  • model: A "luz_module_fitted" object from luz (torch framework). This object can be used to predicting.

  • predictors: A tibble with quantitative (c column names) and qualitative (f column names) variables use for modeling.

  • performance: Averaged performance metrics (see adm_eval).

  • performance_part: Performance metrics for each replica and partition.

  • predicted_part: Observed and predicted abundance for each test partition.

Examples

if (FALSE) { # \dontrun{
require(dplyr)

# Database with species abundance and x and y coordinates
data("sppabund")

# Extract data for a single species
some_sp <- sppabund %>%
  dplyr::filter(species == "Species one") %>%
  dplyr::select(-.part2, -.part3)

# Explore reponse variables
some_sp$ind_ha %>% range()
some_sp$ind_ha %>% hist()

# Here we balance number of absences
some_sp <-
  balance_dataset(some_sp, response = "ind_ha", absence_ratio = 0.2)

# Generate a architecture
dnn_arch <- generate_dnn_architecture(
  number_of_features = 3,
  number_of_outputs = 1,
  number_of_hidden_layers = 3,
  hidden_layers_size = c(8, 16, 8),
  batch_norm = TRUE
)

# Fit a NET model
mdnn <- fit_abund_dnn(
  data = some_sp,
  response = "ind_ha",
  predictors = c("bio12", "elevation", "sand"),
  predictors_f = NULL,
  partition = ".part",
  learning_rate = 0.01,
  n_epochs = 10,
  batch_size = 32,
  validation_patience = 2,
  fitting_patience = 5,
  custom_architecture = dnn_arch,
  verbose = TRUE,
  predict_part = TRUE
)

mdnn
} # }