tiny_dnn
1.0.0
A header only, dependency-free deep learning framework in C++11
|
base class of all kind of NN layers More...
#include <layer.h>
Public Member Functions | |
layer (const std::vector< vector_type > &in_type, const std::vector< vector_type > &out_type) | |
Defaul layer constructor that instantiates a N-input, M-output layer. More... | |
layer (const layer &)=default | |
layer & | operator= (const layer &)=default |
void | set_parallelize (bool parallelize) |
void | set_backend (std::shared_ptr< core::backend > backend) |
void | set_backend_type (core::backend_t backend_type) |
bool | parallelize () const |
core::backend_t | backend_type () const |
core::backend_t | engine () const |
virtual std::string | kernel_file () const |
virtual std::string | kernel_header () const |
virtual void | createOp () |
void | setDevice (const Device &device) |
Device * | device () const |
std::shared_ptr< core::backend > | backend () |
number of incoming edges in this layer | |
serial_size_t | in_channels () const |
number of outgoing edges in this layer | |
serial_size_t | out_channels () const |
serial_size_t | in_data_size () const |
serial_size_t | out_data_size () const |
std::vector< shape3d > | in_data_shape () |
std::vector< shape3d > | out_data_shape () |
serial_size_t | in_size () const |
! More... | |
serial_size_t | out_size () const |
! More... | |
std::vector< const vec_t * > | weights () const |
std::vector< vec_t * > | weights () |
std::vector< tensor_t * > | weights_grads () |
std::vector< edgeptr_t > | inputs () |
std::vector< edgeptr_t > | outputs () |
std::vector< edgeptr_t > | outputs () const |
void | set_out_grads (const std::vector< tensor_t > &grad) |
void | set_in_data (const std::vector< tensor_t > &data) |
std::vector< tensor_t > | output () const |
std::vector< vector_type > | in_types () const |
std::vector< vector_type > | out_types () const |
void | set_trainable (bool trainable) |
bool | trainable () const |
virtual std::pair< float_t, float_t > | out_value_range () const |
return output value range used only for calculating target value from label-id in final(output) layer override properly if the layer is intended to be used as output layer | |
virtual std::vector< shape3d > | in_shape () const =0 |
array of input shapes (width x height x depth) | |
virtual std::vector< shape3d > | out_shape () const =0 |
array of output shapes (width x height x depth) | |
virtual std::string | layer_type () const =0 |
name of layer, should be unique for each concrete class | |
virtual serial_size_t | fan_in_size () const |
number of incoming connections for each output unit used only for weight/bias initialization methods which require fan-in size (e.g. More... | |
virtual serial_size_t | fan_out_size () const |
number of outgoing connections for each input unit used only for weight/bias initialization methods which require fan-out size (e.g. More... | |
template<typename WeightInit > | |
layer & | weight_init (const WeightInit &f) |
template<typename BiasInit > | |
layer & | bias_init (const BiasInit &f) |
template<typename WeightInit > | |
layer & | weight_init (std::shared_ptr< WeightInit > f) |
template<typename BiasInit > | |
layer & | bias_init (std::shared_ptr< BiasInit > f) |
template<typename Archive > | |
void | serialize (Archive &ar) |
virtual void | save (std::ostream &os) const |
virtual void | load (std::istream &is) |
virtual void | load (const std::vector< float_t > &src, int &idx) |
virtual image | output_to_image (size_t channel=0) const |
< visualize latest output of this layer default implementation interpret output as 1d-vector, so "visual" layer(like convolutional layer) should override this for better visualization. | |
virtual void | forward_propagation (const std::vector< tensor_t * > &in_data, std::vector< tensor_t * > &out_data)=0 |
virtual void | back_propagation (const std::vector< tensor_t * > &in_data, const std::vector< tensor_t * > &out_data, std::vector< tensor_t * > &out_grad, std::vector< tensor_t * > &in_grad)=0 |
return delta of previous layer (delta=\frac{dE}{da}, a=wx in fully-connected layer) More... | |
virtual void | post_update () |
return delta2 of previous layer (delta2=\frac{d^2E}{da^2}, diagonal of hessian matrix) it is never called if optimizer is hessian-free | |
virtual void | set_context (net_phase ctx) |
notify changing context (train <=> test) | |
std::vector< tensor_t > | forward (const std::vector< tensor_t > &input) |
std::vector< tensor_t > | backward (const std::vector< tensor_t > &out_grads) |
void | forward () |
void | backward () |
void | setup (bool reset_weight) |
void | init_weight () |
void | clear_grads () |
void | update_weight (optimizer *o, serial_size_t batch_size) |
bool | has_same_weights (const layer &rhs, float_t eps) const |
virtual void | set_sample_count (serial_size_t sample_count) |
template<class Archive > | |
void | serialize_prolog (Archive &ar) |
![]() | |
node (serial_size_t in_size, serial_size_t out_size) | |
const std::vector< edgeptr_t > & | prev () const |
const std::vector< edgeptr_t > & | next () const |
serial_size_t | prev_port (const edge &e) const |
serial_size_t | next_port (const edge &e) const |
std::vector< node * > | prev_nodes () const |
std::vector< node * > | next_nodes () const |
Static Public Member Functions | |
template<typename InputArchive > | |
static std::shared_ptr< layer > | load_layer (InputArchive &ia) |
generate layer from cereal's Archive | |
template<typename OutputArchive > | |
static void | save_layer (OutputArchive &oa, const layer &l) |
Protected Attributes | |
bool | initialized_ |
Flag indication whether the layer/node is initialized. | |
bool | parallelize_ |
Flag indicating whether the layer/node operations ara paralellized. | |
serial_size_t | in_channels_ |
The number of input vectors/edges. | |
serial_size_t | out_channels_ |
The number of output vectors/edges. | |
std::vector< vector_type > | in_type_ |
Vector containing the type of data for inputs. | |
std::vector< vector_type > | out_type_ |
Vector containing the type of data for outputs. | |
core::backend_t | backend_type_ |
The current backend type for operations. | |
std::shared_ptr< core::backend > | backend_ |
The backend instance (deprecated) | |
Device * | device_ptr_ = nullptr |
Pointer to the device on which the layer/node will run. | |
![]() | |
std::vector< edgeptr_t > | prev_ |
std::vector< edgeptr_t > | next_ |
Friends | |
void | connection_mismatch (const layer &from, const layer &to) |
base class of all kind of NN layers
sub-class should override these methods:
|
inline |
Defaul layer constructor that instantiates a N-input, M-output layer.
in_type[N] | type of input vector (data, weight, bias...) |
out_type[M] | type of output vector |
|
pure virtual |
return delta of previous layer (delta=\frac{dE}{da}, a=wx in fully-connected layer)
in_data | input vectors (same vectors as forward_propagation) |
out_data | output vectors (same vectors as forward_propagation) |
out_grad | gradient of output vectors (i-th vector correspond with out_data[i]) |
in_grad | gradient of input vectors (i-th vector correspond with in_data[i]) |
Implemented in tiny_dnn::slice_layer, tiny_dnn::quantized_fully_connected_layer< Activation >, tiny_dnn::quantized_deconvolutional_layer< Activation >, tiny_dnn::quantized_convolutional_layer< Activation >, tiny_dnn::power_layer, tiny_dnn::partial_connected_layer< Activation >, tiny_dnn::partial_connected_layer< activation::identity >, tiny_dnn::max_pooling_layer< Activation >, tiny_dnn::lrn_layer< Activation >, tiny_dnn::linear_layer< Activation >, tiny_dnn::input_layer, tiny_dnn::fully_connected_layer< Activation >, tiny_dnn::dropout_layer, tiny_dnn::deconvolutional_layer< Activation >, tiny_dnn::convolutional_layer< Activation >, tiny_dnn::concat_layer, tiny_dnn::batch_normalization_layer, tiny_dnn::average_unpooling_layer< Activation >, tiny_dnn::average_pooling_layer< Activation >, and tiny_dnn::elementwise_add_layer.
|
inlinevirtual |
number of incoming connections for each output unit used only for weight/bias initialization methods which require fan-in size (e.g.
xavier) override if the layer has trainable weights, and scale of initialization is important
Reimplemented in tiny_dnn::quantized_fully_connected_layer< Activation >, tiny_dnn::quantized_deconvolutional_layer< Activation >, tiny_dnn::quantized_convolutional_layer< Activation >, tiny_dnn::partial_connected_layer< Activation >, tiny_dnn::partial_connected_layer< activation::identity >, tiny_dnn::max_unpooling_layer< Activation >, tiny_dnn::max_pooling_layer< Activation >, tiny_dnn::lrn_layer< Activation >, tiny_dnn::fully_connected_layer< Activation >, tiny_dnn::dropout_layer, tiny_dnn::deconvolutional_layer< Activation >, tiny_dnn::convolutional_layer< Activation >, and tiny_dnn::batch_normalization_layer.
|
inlinevirtual |
number of outgoing connections for each input unit used only for weight/bias initialization methods which require fan-out size (e.g.
xavier) override if the layer has trainable weights, and scale of initialization is important
Reimplemented in tiny_dnn::quantized_fully_connected_layer< Activation >, tiny_dnn::quantized_deconvolutional_layer< Activation >, tiny_dnn::quantized_convolutional_layer< Activation >, tiny_dnn::partial_connected_layer< Activation >, tiny_dnn::partial_connected_layer< activation::identity >, tiny_dnn::max_unpooling_layer< Activation >, tiny_dnn::max_pooling_layer< Activation >, tiny_dnn::lrn_layer< Activation >, tiny_dnn::fully_connected_layer< Activation >, tiny_dnn::dropout_layer, tiny_dnn::deconvolutional_layer< Activation >, tiny_dnn::convolutional_layer< Activation >, and tiny_dnn::batch_normalization_layer.
|
pure virtual |
in_data | input vectors of this layer (data, weight, bias) |
out_data | output vectors |
Implemented in tiny_dnn::slice_layer, tiny_dnn::quantized_fully_connected_layer< Activation >, tiny_dnn::quantized_deconvolutional_layer< Activation >, tiny_dnn::quantized_convolutional_layer< Activation >, tiny_dnn::power_layer, tiny_dnn::partial_connected_layer< Activation >, tiny_dnn::partial_connected_layer< activation::identity >, tiny_dnn::max_pooling_layer< Activation >, tiny_dnn::lrn_layer< Activation >, tiny_dnn::linear_layer< Activation >, tiny_dnn::input_layer, tiny_dnn::fully_connected_layer< Activation >, tiny_dnn::dropout_layer, tiny_dnn::deconvolutional_layer< Activation >, tiny_dnn::convolutional_layer< Activation >, tiny_dnn::concat_layer, tiny_dnn::batch_normalization_layer, tiny_dnn::average_unpooling_layer< Activation >, tiny_dnn::average_pooling_layer< Activation >, and tiny_dnn::elementwise_add_layer.
|
inline |
!
|
inline |
!