A struct containing necessary information for training neural networks. More...

#include <LibtorchArtificialNeuralNetTrainer.h>

Public Attributes
MooseEnum	optimizer_type = MooseEnum("adam=0 adagrad=1 rmsprop=2 sgd=3", "adam")
	The type of optimizer we want to use for training, adam is the default due to its robustness and fast convergence. More...

unsigned int	num_epochs = 1
	Number of iterations we want to perform on the whole dataset. More...

unsigned int	num_batches = 1
	Number of batches we want to split the dataset into. More...

bool	print_loss = false
	If we want to print additional information during training. More...

unsigned int	print_epoch_loss = 1
	The frequency of training loss print to console. More...

Real	rel_loss_tol = 1e-12
	The relative loss tolerance where the training shall stop. More...

Real	learning_rate = 1e-3
	The learning rate for the optimizers. More...

unsigned int	parallel_processes = 1
	The number of allowed parallel processes. More...

bool	allow_duplicates = false
	Parameter for sampling. More...

Detailed Description

A struct containing necessary information for training neural networks.

This is an easy way to pass multiple parameters to the neural networks and follows the design of multiple libtorch classes.

Definition at line 27 of file LibtorchArtificialNeuralNetTrainer.h.

Member Data Documentation

◆ allow_duplicates

bool Moose::LibtorchTrainingOptions::allow_duplicates = false

Parameter for sampling.

This actually allows random samplers to use the same sample multiple times. Sometimes comes in handy when only a few of them are needed to fill up empty slots in the last batch. NOTE: This may introduce a considerable BIAS!

Definition at line 51 of file LibtorchArtificialNeuralNetTrainer.h.

Referenced by Moose::LibtorchArtificialNeuralNetTrainer< SamplerType >::train().

◆ learning_rate

Real Moose::LibtorchTrainingOptions::learning_rate = 1e-3

The learning rate for the optimizers.

Definition at line 43 of file LibtorchArtificialNeuralNetTrainer.h.

Referenced by Moose::LibtorchArtificialNeuralNetTrainer< SamplerType >::createOptimizer().

◆ num_batches

unsigned int Moose::LibtorchTrainingOptions::num_batches = 1

Number of batches we want to split the dataset into.

Definition at line 35 of file LibtorchArtificialNeuralNetTrainer.h.

Referenced by Moose::LibtorchArtificialNeuralNetTrainer< SamplerType >::train().

◆ num_epochs

unsigned int Moose::LibtorchTrainingOptions::num_epochs = 1

Number of iterations we want to perform on the whole dataset.

Definition at line 33 of file LibtorchArtificialNeuralNetTrainer.h.

◆ optimizer_type

MooseEnum Moose::LibtorchTrainingOptions::optimizer_type = MooseEnum("adam=0 adagrad=1 rmsprop=2 sgd=3", "adam")

The type of optimizer we want to use for training, adam is the default due to its robustness and fast convergence.

Definition at line 31 of file LibtorchArtificialNeuralNetTrainer.h.

Referenced by Moose::LibtorchArtificialNeuralNetTrainer< SamplerType >::createOptimizer().

◆ parallel_processes

unsigned int Moose::LibtorchTrainingOptions::parallel_processes = 1

The number of allowed parallel processes.

It is necessary to bound it because the user may want to run simulations with more processors than data samples which leads to segmentation faults is a process remains with no sample

Definition at line 47 of file LibtorchArtificialNeuralNetTrainer.h.

Referenced by Moose::LibtorchArtificialNeuralNetTrainer< SamplerType >::train().