train_tower_model: Train tower model with keras

In treysp/coolit.train: Train model to identify water cooling towers from aerial orthoimagery

Description

Train tower model with keras

Usage

train_tower_model(img_dir = "data/tiles/splits/orig",
  output_dir = "output", img_size = c(50, 50),
  img_horizontal_flip = TRUE, img_vertical_flip = TRUE,
  batch_size = 32, base_model = "vgg16", save_best_model_only = TRUE,
  dense_structure = list(list(units = 256, dropout = 0.2), list(units =
  128, dropout = 0.2)), dense_optimizer = "rmsprop", dense_lr = 1e-05,
  dense_steps_per_epoch = 100, dense_epochs = 30,
  dense_validation_steps = 50, first_ft_unfreeze = "block4_conv1",
  first_ft_optimizer = "rmsprop", first_ft_lr = 1e-05,
  first_ft_steps_per_epoch = 100, first_ft_epochs = 30,
  first_ft_validation_steps = 50, do_second_ft = TRUE,
  second_ft_unfreeze = "block3_conv1", second_ft_optimizer = "rmsprop",
  second_ft_lr = 5e-06, second_ft_steps_per_epoch = 100,
  second_ft_epochs = 30, second_ft_validation_steps = 50,
  class_weights = NULL)

Arguments

img_dir	Directory containing properly structured 'train' and 'validation' directories
output_dir	Directory where model outputs should be save
img_size	Numeric vector length 2 containing the training image size in pixels (order row, column)
img_horizontal_flip	Image augmentation: should images be flipped horizontally during training?
img_vertical_flip	Image augmentation: should images be flipped vertically during training?
batch_size	Training model batch size (single number).
base_model	Character containing the name of a base model available in the R 'keras' package via an 'application_[base model name]' function. For example, parameter value "vgg16" would call the keras 'application_vgg16' function.
save_best_model_only	Should training callback save model at every epoch or only retain model with the best validation loss?
dense_structure	List of lists specifying structure of dense layers added to base model. See examples.
dense_optimizer	Character containing the name of an optimizer available in the R 'keras' package via an 'optimizer_[optimizer name]' function. For example, parameter value "rmsprop" would call the keras 'optimizer_rmsprop' function.
dense_lr	Learning rate for dense layer training (single number).
dense_steps_per_epoch	Steps per epoch for dense layer training (single number).
dense_epochs	Number of epochs for dense layer training (single number).
dense_validation_steps	Number of validation steps for dense layer training (single number).
first_ft_unfreeze	Name of the base model layer where weights should be unfrozen for first fine-tune training. Must be a valid layer name for the base model specified in the 'base_model' parameter.
first_ft_optimizer	Character containing the name of an optimizer available in the R 'keras' package via an 'optimizer_[optimizer name]' function.
first_ft_lr	Learning rate for first fine-tune training (single number).
first_ft_steps_per_epoch	Steps per epoch for first fine-tune training (single number).
first_ft_epochs	Number of epochs for first fine-tune training (single number).
first_ft_validation_steps	Number of validation steps for first fine-tune training (single number).
do_second_ft	Do a second fine-tune training (TRUE or FALSE)?
second_ft_unfreeze	Name of the base model layer where weights should be unfrozen for second fine-tune training. Must be a valid layer name for the base model specified in the 'base_model' parameter.
second_ft_optimizer	Character containing the name of an optimizer available in the R 'keras' package via an 'optimizer_[optimizer name]' function.
second_ft_lr	Learning rate for second fine-tune training (single number).
second_ft_steps_per_epoch	Steps per epoch for second fine-tune training (single number).
second_ft_epochs	Number of epochs for second fine-tune training (single number).
second_ft_validation_steps	Number of validation steps for second fine-tune training (single number).
class_weights	Named list with weights corresponding to outcome classes - see examples. Optional - default is weights inversely proportional to outcome class distribution in training set.

treysp/coolit.train documentation built on Oct. 10, 2019, 3:24 p.m.