the Neural_Network More...

#include <network.h>

Collaboration diagram for bc::nn::NeuralNetwork< Layers >:

Public Types
using	self = NeuralNetwork< Layers... >

using	is_recurrent = bc::traits::any< detail::is_recurrent_layer, Layers... >

using	layer_chain = LayerChain< bc::traits::Integer< 0 >, void, Layers... >

Public Member Functions
	NeuralNetwork (Layers... layers)
	Basic Constructor for Neural Networks. More...

template<class T >
auto	forward_propagation (const T &tensor)
	Calls forward propagation on each of the neural_network's layers The time index will be set to zero prior to each call forward_call. More...

template<class T >
auto	back_propagation (const T &tensor)
	Calls back-propagation on each of the neural_network's layers The time index will be incremented after each call to back_propagation. More...

template<class T >
auto	predict (const T &tensor)
	Returns the output of a single batch. More...

template<class T >
auto	single_predict (const T &tensor)
	Returns the output of a single input. More...

template<int X>
auto &	get_layer () const
	Returns a const reference to the layer specified by the given index. More...

template<int X>
auto &	get_layer ()
	Returns a reference to the layer specified by the given index. More...

void	set_learning_rate (double lr)
	Sets the learning for each layer in the Neural_Network. More...

double	get_learning_rate () const
	Returns the current global learning rate. More...

void	set_batch_size (int batch_sz)
	Sets the batch_size of the entire Neural_Network The intermediate values are discarded when setting the batch_size. More...

void	copy_training_data_to_single_predict (int batch_index)
	Copies the given batch_index from the training cell_state to the inference cell_state. More...

void	zero_time_index ()
	Sets the expression_template time_index to zero. More...

void	update_weights ()
	Updates the weights of each Layer based upon the current stored gradients. More...

bc::size_t	input_size () const
	returns the input_size of the first layer More...

bc::size_t	output_size () const
	returns the output_size of the last layer More...

bc::size_t	batch_size () const
	returns the batch_size of the neural network. More...

std::string	get_string_architecture () const
	Returns a yaml representation of the neural network. More...

void	save (std::string directory_name)
	Creates the directory `directory_name` using mkdir. More...

void	load (std::string directory_name)
	Loads a neural network from a previously saved instance. More...

Public Attributes
layer_chain	m_layer_chain

double	m_learning_rate = m_layer_chain.head().layer().get_learning_rate()

bc::size_t	m_batch_size = 1

Detailed Description

template<class ... Layers>
struct bc::nn::NeuralNetwork< Layers >

the Neural_Network

Member Typedef Documentation

◆ is_recurrent

template<class ... Layers>

using bc::nn::NeuralNetwork< Layers >::is_recurrent = bc::traits::any<detail::is_recurrent_layer, Layers...>

◆ layer_chain

template<class ... Layers>

using bc::nn::NeuralNetwork< Layers >::layer_chain = LayerChain< bc::traits::Integer<0>, void, Layers...>

◆ self

template<class ... Layers>

using bc::nn::NeuralNetwork< Layers >::self = NeuralNetwork<Layers...>

Constructor & Destructor Documentation

◆ NeuralNetwork()

template<class ... Layers>

bc::nn::NeuralNetwork< Layers >::NeuralNetwork ( Layers... layers )

inline

Basic Constructor for Neural Networks.

Accepts a variadic parameter pack of Layer-like objects.

Member Function Documentation

◆ back_propagation()

template<class ... Layers>

template<class T >

auto bc::nn::NeuralNetwork< Layers >::back_propagation ( const T & tensor )

inline

Calls back-propagation on each of the neural_network's layers The time index will be incremented after each call to back_propagation.

Returns: the error of the inputs at the current time_index.

◆ batch_size()

template<class ... Layers>

bc::size_t bc::nn::NeuralNetwork< Layers >::batch_size ( ) const

inline

returns the batch_size of the neural network.

◆ copy_training_data_to_single_predict()

template<class ... Layers>

void bc::nn::NeuralNetwork< Layers >::copy_training_data_to_single_predict ( int batch_index )

inline

Copies the given batch_index from the training cell_state to the inference cell_state.

The time_index is set to zero after copying.

◆ forward_propagation()

template<class ... Layers>

template<class T >

auto bc::nn::NeuralNetwork< Layers >::forward_propagation ( const T & tensor )

inline

Calls forward propagation on each of the neural_network's layers The time index will be set to zero prior to each call forward_call.

Arguments: A Tensor type with the same shape as the first layer's return value of 'get_batched_input_shape()'

Returns: If the neural-network terminates with an Output_Layer forward-propagation will return a 'shallow-copy' of the output.

A shallow-copy will have the behavior of a Tensor class but will act as a handle to the underlying Tensor. (Passing by value will not incur a copy).

◆ get_layer() [1/2]

template<class ... Layers>

template<int X>

auto& bc::nn::NeuralNetwork< Layers >::get_layer ( ) const

inline

Returns a const reference to the layer specified by the given index.

◆ get_layer() [2/2]

template<class ... Layers>

template<int X>

auto& bc::nn::NeuralNetwork< Layers >::get_layer ( )

inline

Returns a reference to the layer specified by the given index.

◆ get_learning_rate()

template<class ... Layers>

double bc::nn::NeuralNetwork< Layers >::get_learning_rate ( ) const

inline

Returns the current global learning rate.

◆ get_string_architecture()

template<class ... Layers>

std::string bc::nn::NeuralNetwork< Layers >::get_string_architecture ( ) const

inline

Returns a yaml representation of the neural network.

◆ input_size()

template<class ... Layers>

bc::size_t bc::nn::NeuralNetwork< Layers >::input_size ( ) const

inline

returns the input_size of the first layer

◆ load()

template<class ... Layers>

void bc::nn::NeuralNetwork< Layers >::load ( std::string directory_name )

inline

Loads a neural network from a previously saved instance.

Load expects the neural-network to have been unused in the previous state.

◆ output_size()

template<class ... Layers>

bc::size_t bc::nn::NeuralNetwork< Layers >::output_size ( ) const

inline

returns the output_size of the last layer

◆ predict()

template<class ... Layers>

template<class T >

auto bc::nn::NeuralNetwork< Layers >::predict ( const T & tensor )

inline

Returns the output of a single batch.

Predict is identical to forward propagation except it does not cache the intermediate values required for back_propagation. The inputs and outputs of each layer however are cached, as they are required for most recurrent neural networks.

◆ save()

template<class ... Layers>

void bc::nn::NeuralNetwork< Layers >::save ( std::string directory_name )

inline

Creates the directory directory_name using mkdir.

Than outputs an architecture.yaml file with a description of the neural network. Than iterates through each layer and calls 'save(int layer_index, std::string directory_name)'

◆ set_batch_size()

template<class ... Layers>

void bc::nn::NeuralNetwork< Layers >::set_batch_size ( int batch_sz )

inline

Sets the batch_size of the entire Neural_Network The intermediate values are discarded when setting the batch_size.

◆ set_learning_rate()

template<class ... Layers>

void bc::nn::NeuralNetwork< Layers >::set_learning_rate ( double lr )

inline

Sets the learning for each layer in the Neural_Network.

Individual layer's learning rates can be set by accessing them via 'get_layer<int index>()' and calling 'set_learning_rate'

◆ single_predict()

template<class ... Layers>

template<class T >

auto bc::nn::NeuralNetwork< Layers >::single_predict ( const T & tensor )

inline

Returns the output of a single input.

single_predict is identical to predict except it accepts a non-batched input to forward-propagate on.

◆ update_weights()

template<class ... Layers>

void bc::nn::NeuralNetwork< Layers >::update_weights ( )

inline

Updates the weights of each Layer based upon the current stored gradients.

◆ zero_time_index()

template<class ... Layers>

void bc::nn::NeuralNetwork< Layers >::zero_time_index ( )

inline

Sets the expression_template time_index to zero.

The time_index dictates what intermediates to return during back_propagation and forward_propagation of a recurrent neural network.

This value is not relevant in non-recurrent neural networks.

Member Data Documentation

◆ m_batch_size

template<class ... Layers>

bc::size_t bc::nn::NeuralNetwork< Layers >::m_batch_size = 1

◆ m_layer_chain

template<class ... Layers>

layer_chain bc::nn::NeuralNetwork< Layers >::m_layer_chain

◆ m_learning_rate

template<class ... Layers>

double bc::nn::NeuralNetwork< Layers >::m_learning_rate = m_layer_chain.head().layer().get_learning_rate()

The documentation for this struct was generated from the following file:

blackcat/neural_networks/network.h

Public Types

Public Member Functions

Public Attributes

Detailed Description

template<class ... Layers> struct bc::nn::NeuralNetwork< Layers >

Member Typedef Documentation

◆ is_recurrent

◆ layer_chain

◆ self

Constructor & Destructor Documentation

◆ NeuralNetwork()

Member Function Documentation

◆ back_propagation()

◆ batch_size()

◆ copy_training_data_to_single_predict()

◆ forward_propagation()

◆ get_layer() [1/2]

◆ get_layer() [2/2]

◆ get_learning_rate()

◆ get_string_architecture()

◆ input_size()

◆ load()

◆ output_size()

◆ predict()

◆ save()

◆ set_batch_size()

◆ set_learning_rate()

◆ single_predict()

◆ update_weights()

◆ zero_time_index()

Member Data Documentation

◆ m_batch_size

◆ m_layer_chain

◆ m_learning_rate

template<class ... Layers>
struct bc::nn::NeuralNetwork< Layers >