LBANN  0.103.0
LivermoreBigArtificialNeuralNetworkToolkit
lbann::base_convolution_layer< TensorDataType, Device > Class Template Referenceabstract

Computation kernels for convolution and deconvolution layers. More...

#include <base_convolution.hpp>

Inheritance diagram for lbann::base_convolution_layer< TensorDataType, Device >:
[legend]
Collaboration diagram for lbann::base_convolution_layer< TensorDataType, Device >:
[legend]

Public Types

Public Types
using WeightsType = data_type_weights< TensorDataType >
 The concrete weights type used by this object. More...
 
using OptimizerType = data_type_optimizer< TensorDataType >
 The concrete optimizer type used by this object. More...
 
template<El::Device D>
using DMatDT = El::Matrix< TensorDataType, D >
 
using ScalingType = TensorDataType
 
- Public Types inherited from lbann::data_type_layer< TensorDataType >
using InputAbsDistMatrixType = El::AbstractDistMatrix< TensorDataType >
 The tensor type expected in this object. More...
 
using OutputAbsDistMatrixType = El::AbstractDistMatrix< TensorDataType >
 
using InputAbsDistMatReadProxyType = El::AbstractDistMatrixReadDeviceProxy< TensorDataType, D >
 The proxy tensor type expected in this object. More...
 
using OutputAbsDistMatReadProxyType = El::AbstractDistMatrixReadDeviceProxy< TensorDataType, D >
 
using InputAbsMatrixType = El::AbstractMatrix< TensorDataType >
 The local tensor type expected in this object. More...
 
using OutputAbsMatrixType = El::AbstractMatrix< TensorDataType >
 
using WeightsProxyType = weights_proxy< TensorDataType >
 The proxy type for weights used by this object. More...
 

Public Member Functions

const std::vector< int > & get_conv_dims () const
 Get convolutional layer parameters. More...
 
const std::vector< int > & get_pads () const
 
const std::vector< int > & get_strides () const
 
const std::vector< int > & get_dilations () const
 
 base_convolution_layer (int num_data_dims, int output_channels, std::vector< int > conv_dims, std::vector< int > pads, std::vector< int > strides, std::vector< int > dilations, int groups, bool has_bias)
 
 base_convolution_layer (const base_convolution_layer &other)
 
base_convolution_layeroperator= (const base_convolution_layer &other)
 
 ~base_convolution_layer ()
 
description get_description () const override
 Human-readable description. More...
 
void setup_dims () override
 Setup tensor dimensions Called by the 'setup' function. If there are any input tensors, the base method sets all uninitialized output tensor dimensions equal to the first input tensor dimensions. More...
 
void setup_data (size_t max_mini_batch_size) override
 Setup layer data. The kernel weights are setup in the convolution and deconvolution classes. More...
 
void setup_gpu () override
 Initialize GPU objects. More...
 
Serialization
template<typename ArchiveT >
void serialize (ArchiveT &ar)
 
- Public Member Functions inherited from lbann::data_type_layer< TensorDataType >
 data_type_layer (lbann_comm *, bool persistent_error_signals=false)
 
virtual ~data_type_layer ()=default
 
std::string get_datatype_name () const override
 
void forward_prop () final
 
void summarize_matrices (lbann_summary &summarizer, int step) override
 
void check_setup () override
 
const OutputAbsDistMatrixTypeget_activations (const Layer &child) const override
 
OutputAbsDistMatrixTypeget_activations (int child_index=0)
 
const OutputAbsDistMatrixTypeget_activations (int child_index=0) const
 
const InputAbsDistMatrixTypeget_error_signals (const Layer &parent) const override
 
InputAbsDistMatrixTypeget_error_signals (int parent_index=0)
 
const InputAbsDistMatrixTypeget_error_signals (int parent_index=0) const
 
El::Int current_output_mini_batch_size () const override
 
El::Int infer_mini_batch_size_from_parents_or_default_to_current () const override
 
OutputAbsDistMatrixTypeget_temp_grad ()
 
InputAbsDistMatrixTypeget_branch_tag_input (int tag)
 
std::vector< std::unique_ptr< InputAbsDistMatrixType > > & get_branch_tag_input_vector ()
 
std::vector< std::unique_ptr< OutputAbsDistMatrixType > > & get_all_activations ()
 
std::vector< std::unique_ptr< InputAbsDistMatrixType > > & get_all_prev_activations ()
 
std::vector< std::unique_ptr< OutputAbsDistMatrixType > > & get_all_prev_error_signals ()
 
std::vector< std::unique_ptr< InputAbsDistMatrixType > > & get_all_error_signals ()
 
OutputAbsMatrixTypeget_local_activations (int child_index=0)
 
const OutputAbsMatrixTypeget_local_activations (int child_index=0) const
 
InputAbsMatrixTypeget_local_error_signals (int parent_index=0)
 
const InputAbsMatrixTypeget_local_error_signals (int parent_index=0) const
 
void set_keep_error_signals (bool) override
 Set whether to keep or dynamically reallocate error signals. More...
 
El::mpi::Comm & get_subgrid_comm ()
 
void serialize (ArchiveT &ar)
 
- Public Member Functions inherited from lbann::Layer
void write_proto (lbann_data::Layer &proto) const
 Write layer to proto file. More...
 
lbann_commget_comm () const
 
int get_grid_tag () const noexcept
 Identifying tag for process grid. More...
 
void set_grid_tag (int tag)
 Set process grid. More...
 
bool runs_inplace () const
 If true, the layer will run in-place (the input and output activations point to the same tensor). Value is set during graph setup (in setup_pointers) based on layer traits and neighboring layers. More...
 
bool distconv_enabled () const
 Indicate whether distconv is enabled. More...
 
 Layer ()
 
virtual ~Layer ()=default
 
virtual Layercopy () const =0
 Copy function. This function dynamically allocates memory for a layer instance and instantiates a copy. The caller is responsible for deallocating the instance. More...
 
void set_name (const std::string name)
 Set the layer instance's name. Each layer in a model should have a unique, preferably human-readable, name. More...
 
void set_model (model *m)
 Set the model that manages this layer. More...
 
std::string get_name () const
 Get the layer instance's name. More...
 
modelget_model () const noexcept
 Get a reference to the model that manages this layer. More...
 
virtual std::string get_type () const =0
 Get the layer type's name. More...
 
virtual data_layout get_data_layout () const =0
 Get data layout of the data tensors. We assume that the data layouts of the previous activations, activations, previous error signals, and error signals are the same. Each concrete layer that is templated on its data layout should override this function to return its template parameter. More...
 
virtual El::Device get_device_allocation () const =0
 Get the device allocation for the data tensors. We assume that the decice allocation of the previous activations, activations, previous error signals, and error signals are the same. Each concrete layer that is templated on its device allocation should override this function to return its template parameter. More...
 
int get_expected_num_parent_layers () const noexcept
 Get expected number of parent layers. A negative value indicates no limit. More...
 
int get_expected_num_child_layers () const noexcept
 Get expected number of child layers. A negative value indicates no limit. More...
 
virtual int get_backprop_requirements () const
 Returns the necessary tensors for computing backpropagation. More...
 
ParallelStrategyget_parallel_strategy () noexcept
 Get the parallel strategy for the layer. More...
 
ParallelStrategy const & get_parallel_strategy () const noexcept
 Get the parallel strategy for the layer. More...
 
virtual bool can_run_inplace () const
 If True, the computation can run in-place (feeding each input activations tensor as the corresponding output activations) More...
 
bool using_gpus () const noexcept
 Whether the layer is using a GPU implementation. More...
 
void back_prop ()
 Backward propagation step. Given the objective function gradients w.r.t. the output tensors, compute the gradients w.r.t. the input tensors and w.r.t. the weights. This is essentially an application of the chain rule. More...
 
bool update ()
 Update step. Update the layer's internal members. Note that the optimization step for the weights happens elsewhere. More...
 
virtual void setup (size_t max_mini_batch_size, const std::vector< El::Grid *> &grids)
 Setup layer members. More...
 
void summarize_stats (lbann_summary &summarizer, int step)
 
void reset_counters ()
 Reset layer stat counters. More...
 
void set_communication_flag (SubGraphCommunication type)
 
SubGraphCommunication get_communication_flag ()
 
void set_num_spliting_groups (El::Int spliting_groups)
 
El::Int get_num_spliting_groups () const
 
std::shared_ptr< El::Grid > get_mygrid () const
 
void reset_inter_subgrid_vc_comm (std::shared_ptr< El::mpi::Comm > mpi_comm)
 
void set_subgraph_parallelism_execution ()
 
bool subgraph_parallelism_execution () const noexcept
 
void set_run_layer_in_subgraph ()
 
bool get_run_layer_in_subgraph () const noexcept
 
const Layerget_parent_layer (size_t index=0) const
 
const Layerget_child_layer (size_t index=0) const
 
std::vector< const Layer * > get_parent_layers () const
 
std::vector< const Layer * > get_child_layers () const
 
size_t find_parent_layer_index (const Layer &l) const
 
size_t find_child_layer_index (const Layer &l) const
 
int get_num_parents () const noexcept
 Get number of parent layers. More...
 
int get_num_children () const noexcept
 Get number of child layers. More...
 
void add_parent_layer (ViewingLayerPtr parent)
 Add a parent layer. More...
 
void add_child_layer (ViewingLayerPtr child)
 Add a child layer. More...
 
void replace_parent_layer (ViewingLayerPtr l, size_t index)
 
void replace_child_layer (ViewingLayerPtr l, size_t index)
 
void clear_parent_layers ()
 Remove pointers to parent layers. More...
 
void clear_child_layers ()
 Remove pointers to child layers. More...
 
ViewingLayerPtr get_parent_layer_pointer (size_t index) const
 
ViewingLayerPtr get_child_layer_pointer (size_t index) const
 
virtual std::vector< ViewingLayerPtrget_layer_pointers ()
 List of pointers to other layers. More...
 
virtual void set_layer_pointers (std::vector< ViewingLayerPtr > layers)
 Set list of pointers to other layers. More...
 
std::vector< ViewingWeightsPtrget_weights_pointers () const
 List of pointers to weights. More...
 
void set_weights_pointers (std::vector< ViewingWeightsPtr > ptrs)
 Set list of pointers to weights. More...
 
void replace_weights (Layer const &other_layer)
 Replace weights with another Layer's weights. More...
 
std::vector< int > get_input_dims (size_t input_index=0) const
 Get input tensor dimensions. More...
 
int get_input_size (size_t input_index=0) const
 Get input tensor size. More...
 
std::vector< int > get_output_dims (size_t output_index=0) const
 Get output tensor dimensions. More...
 
int get_output_size (size_t output_index=0) const
 Get output tensor size. More...
 
void set_output_dims (std::vector< int > dims, size_t output_index=0)
 Set output tensor dimensions. More...
 
El::Int infer_mini_batch_size_from_parents () const
 
void set_hint_layer (ViewingLayerPtr l)
 Set hint layer. More...
 
const Layerget_hint_layer () const
 Get hint layer. More...
 
void freeze ()
 
void unfreeze ()
 
bool is_frozen () const
 
template<typename ArchiveT >
void serialize (ArchiveT &ar)
 

Protected Member Functions

virtual std::vector< int > get_kernel_dims () const =0
 
void apply_convolution_dnn (bool during_forward_prop)
 
void apply_transposed_convolution_dnn (bool during_forward_prop)
 
void apply_bias_dnn ()
 
void compute_gradients_dnn (bool using_transposed_convolution)
 
void apply_convolution_im2col (bool during_forward_prop)
 
void apply_transposed_convolution_im2col (bool during_forward_prop)
 
void apply_bias_cpu ()
 
void compute_gradients_im2col (bool using_transposed_convolution)
 
- Protected Member Functions inherited from lbann::data_type_layer< TensorDataType >
InputAbsDistMatrixTypeget_prev_activations (int parent_index=0)
 
const InputAbsDistMatrixTypeget_prev_activations (int parent_index=0) const
 
OutputAbsDistMatrixTypeget_prev_error_signals (int child_index=0)
 
const OutputAbsDistMatrixTypeget_prev_error_signals (int child_index=0) const
 
const InputAbsMatrixTypeget_local_prev_activations (int parent_index=0) const
 
const OutputAbsMatrixTypeget_local_prev_error_signals (int child_index=0) const
 
void setup_matrices (const std::vector< El::Grid * > &grids) override
 
void setup_data (size_t max_mini_batch_size) override
 
void fp_setup_inputs () override
 
void fp_setup_outputs () override
 
void bp_setup_gradient_wrt_inputs () override
 
void bp_compute () override
 
InputAbsDistMatrixType const & weights_values (size_t idx) const
 Get the values matrix for a specific weights object. More...
 
weightsmaster_weights (size_t idx)
 Get a specific master weights object. More...
 
weights const & master_weights (size_t idx) const
 
 data_type_layer (data_type_layer &&other)=default
 Protected lifecycle functions. More...
 
 data_type_layer (data_type_layer const &other)
 
data_type_layeroperator= (data_type_layer &&other)=default
 
data_type_layeroperator= (data_type_layer const &other)
 
- Protected Member Functions inherited from lbann::Layer
void setup_grid ()
 Setup process grid. More...
 
virtual void setup_pointers ()
 Setup layer pointers. Called by the 'setup' function. Pointers to parent/child layers are assumed to be already initialized. More...
 
virtual void fp_compute ()=0
 Apply layer operation. Called by the 'forward_prop' function. Given the input tensors, the output tensors are populated with computed values. More...
 
virtual bool update_compute ()
 Perform the computation for the update step. Returns false if the layer must reset for a new training epoch. More...
 
 Layer (Layer &&other)=default
 
 Layer (Layer const &other)
 
Layeroperator= (Layer &&other)=default
 
Layeroperator= (Layer const &other)
 
void add_weights (ViewingWeightsPtr w)
 
size_t num_weights () const noexcept
 
bool has_weights () const noexcept
 
bool has_weights (size_t idx) const noexcept
 
void set_num_weights (size_t n)
 
void set_weights (size_t idx, ViewingWeightsPtr w)
 
weights const & get_weights (size_t idx) const
 
weightsget_weights (size_t idx)
 
void add_as_gradient_source ()
 
void remove_as_gradient_source ()
 

Protected Attributes

int m_output_channels
 
std::vector< int > m_conv_dims
 Spatial dimensions for convolution kernel. More...
 
std::vector< int > m_pads
 
std::vector< int > m_strides
 
std::vector< int > m_dilations
 
int m_groups
 
ScalingType m_bias_scaling_factor
 
- Protected Attributes inherited from lbann::Layer
int m_expected_num_parent_layers = 1
 
int m_expected_num_child_layers = 1
 Expected number of child layers. A negative value indicates no limit. More...
 
modelm_model = nullptr
 Reference to model managing this layer. More...
 
bool m_frozen
 Avoid back prop if frozen. More...
 
EvalType m_fp_time
 Time spent in forward propagation. More...
 
EvalType m_fp_compute_time
 Time spent in the forward propagation computation. More...
 
EvalType m_bp_time
 Time spent in backward propagation. More...
 
EvalType m_bp_compute_time
 Time spent in the backward propagation computation. More...
 
EvalType m_update_time
 Time spent in updates. More...
 
std::string m_name
 Layer instance's name. Each layer in a model should have a unique, preferably human-readable, name. More...
 
bool m_runs_inplace = false
 If true, the layer will run in-place (the input and output activations point to the same tensor). Value is set during graph setup (in setup_pointers) based on layer traits and neighboring layers. More...
 
int m_grid_tag = -1
 Identifying tag for process grid. More...
 
SubGraphCommunication subgraph_communication_method = PT2PT
 
bool m_subgraph_parallelism_execution = false
 
bool run_layer_in_subgraph = false
 
std::unique_ptr< std::set< int > > m_subgrid_ranks
 
El::Int m_num_spliting_groups = 1
 
std::shared_ptr< El::mpi::Comm > m_interSubGridVCComm
 

Detailed Description

template<typename TensorDataType, El::Device Device>
class lbann::base_convolution_layer< TensorDataType, Device >

Computation kernels for convolution and deconvolution layers.

Definition at line 94 of file base_convolution.hpp.

Member Typedef Documentation

◆ DMatDT

template<typename TensorDataType , El::Device Device>
template<El::Device D>
using lbann::base_convolution_layer< TensorDataType, Device >::DMatDT = El::Matrix<TensorDataType, D>

Definition at line 107 of file base_convolution.hpp.

◆ OptimizerType

template<typename TensorDataType , El::Device Device>
using lbann::base_convolution_layer< TensorDataType, Device >::OptimizerType = data_type_optimizer<TensorDataType>

The concrete optimizer type used by this object.

Definition at line 104 of file base_convolution.hpp.

◆ ScalingType

template<typename TensorDataType , El::Device Device>
using lbann::base_convolution_layer< TensorDataType, Device >::ScalingType = TensorDataType

Definition at line 112 of file base_convolution.hpp.

◆ WeightsType

template<typename TensorDataType , El::Device Device>
using lbann::base_convolution_layer< TensorDataType, Device >::WeightsType = data_type_weights<TensorDataType>

The concrete weights type used by this object.

Definition at line 101 of file base_convolution.hpp.

Constructor & Destructor Documentation

◆ base_convolution_layer() [1/2]

template<typename TensorDataType , El::Device Device>
lbann::base_convolution_layer< TensorDataType, Device >::base_convolution_layer ( int  num_data_dims,
int  output_channels,
std::vector< int >  conv_dims,
std::vector< int >  pads,
std::vector< int >  strides,
std::vector< int >  dilations,
int  groups,
bool  has_bias 
)
Todo:
Remove num_data_dims from arg list

◆ base_convolution_layer() [2/2]

template<typename TensorDataType , El::Device Device>
lbann::base_convolution_layer< TensorDataType, Device >::base_convolution_layer ( const base_convolution_layer< TensorDataType, Device > &  other)

◆ ~base_convolution_layer()

template<typename TensorDataType , El::Device Device>
lbann::base_convolution_layer< TensorDataType, Device >::~base_convolution_layer ( )

Member Function Documentation

◆ apply_bias_cpu()

template<typename TensorDataType , El::Device Device>
void lbann::base_convolution_layer< TensorDataType, Device >::apply_bias_cpu ( )
protected

◆ apply_bias_dnn()

template<typename TensorDataType , El::Device Device>
void lbann::base_convolution_layer< TensorDataType, Device >::apply_bias_dnn ( )
protected

◆ apply_convolution_dnn()

template<typename TensorDataType , El::Device Device>
void lbann::base_convolution_layer< TensorDataType, Device >::apply_convolution_dnn ( bool  during_forward_prop)
protected

Convolution with DNN library.

◆ apply_convolution_im2col()

template<typename TensorDataType , El::Device Device>
void lbann::base_convolution_layer< TensorDataType, Device >::apply_convolution_im2col ( bool  during_forward_prop)
protected

Convolution with im2col GEMM algorithm.

◆ apply_transposed_convolution_dnn()

template<typename TensorDataType , El::Device Device>
void lbann::base_convolution_layer< TensorDataType, Device >::apply_transposed_convolution_dnn ( bool  during_forward_prop)
protected

Transposed convolution with DNN library.

◆ apply_transposed_convolution_im2col()

template<typename TensorDataType , El::Device Device>
void lbann::base_convolution_layer< TensorDataType, Device >::apply_transposed_convolution_im2col ( bool  during_forward_prop)
protected

Transposed convolution with im2col GEMM algorithm.

◆ compute_gradients_dnn()

template<typename TensorDataType , El::Device Device>
void lbann::base_convolution_layer< TensorDataType, Device >::compute_gradients_dnn ( bool  using_transposed_convolution)
protected

◆ compute_gradients_im2col()

template<typename TensorDataType , El::Device Device>
void lbann::base_convolution_layer< TensorDataType, Device >::compute_gradients_im2col ( bool  using_transposed_convolution)
protected

◆ get_conv_dims()

template<typename TensorDataType , El::Device Device>
const std::vector<int>& lbann::base_convolution_layer< TensorDataType, Device >::get_conv_dims ( ) const
inline

Get convolutional layer parameters.

Definition at line 118 of file base_convolution.hpp.

◆ get_description()

template<typename TensorDataType , El::Device Device>
description lbann::base_convolution_layer< TensorDataType, Device >::get_description ( ) const
overridevirtual

Human-readable description.

Reimplemented from lbann::Layer.

◆ get_dilations()

template<typename TensorDataType , El::Device Device>
const std::vector<int>& lbann::base_convolution_layer< TensorDataType, Device >::get_dilations ( ) const
inline

Definition at line 121 of file base_convolution.hpp.

◆ get_kernel_dims()

template<typename TensorDataType , El::Device Device>
virtual std::vector<int> lbann::base_convolution_layer< TensorDataType, Device >::get_kernel_dims ( ) const
protectedpure virtual

◆ get_pads()

template<typename TensorDataType , El::Device Device>
const std::vector<int>& lbann::base_convolution_layer< TensorDataType, Device >::get_pads ( ) const
inline

Definition at line 119 of file base_convolution.hpp.

◆ get_strides()

template<typename TensorDataType , El::Device Device>
const std::vector<int>& lbann::base_convolution_layer< TensorDataType, Device >::get_strides ( ) const
inline

Definition at line 120 of file base_convolution.hpp.

◆ operator=()

template<typename TensorDataType , El::Device Device>
base_convolution_layer& lbann::base_convolution_layer< TensorDataType, Device >::operator= ( const base_convolution_layer< TensorDataType, Device > &  other)

◆ serialize()

template<typename TensorDataType , El::Device Device>
template<typename ArchiveT >
void lbann::base_convolution_layer< TensorDataType, Device >::serialize ( ArchiveT &  ar)

◆ setup_data()

template<typename TensorDataType , El::Device Device>
void lbann::base_convolution_layer< TensorDataType, Device >::setup_data ( size_t  max_mini_batch_size)
overridevirtual

Setup layer data. The kernel weights are setup in the convolution and deconvolution classes.

Reimplemented from lbann::Layer.

◆ setup_dims()

template<typename TensorDataType , El::Device Device>
void lbann::base_convolution_layer< TensorDataType, Device >::setup_dims ( )
overridevirtual

Setup tensor dimensions Called by the 'setup' function. If there are any input tensors, the base method sets all uninitialized output tensor dimensions equal to the first input tensor dimensions.

Reimplemented from lbann::Layer.

Reimplemented in lbann::convolution_layer< TensorDataType, Layout, Device >, and lbann::deconvolution_layer< TensorDataType, Layout, Device >.

◆ setup_gpu()

template<typename TensorDataType , El::Device Device>
void lbann::base_convolution_layer< TensorDataType, Device >::setup_gpu ( )
overridevirtual

Initialize GPU objects.

Reimplemented from lbann::Layer.

Member Data Documentation

◆ m_bias_scaling_factor

template<typename TensorDataType , El::Device Device>
ScalingType lbann::base_convolution_layer< TensorDataType, Device >::m_bias_scaling_factor
protected

Scaling factor for bias term. If the scaling factor is zero, bias is not applied.

Definition at line 145 of file base_convolution.hpp.

◆ m_conv_dims

template<typename TensorDataType , El::Device Device>
std::vector<int> lbann::base_convolution_layer< TensorDataType, Device >::m_conv_dims
protected

Spatial dimensions for convolution kernel.

Excludes number of input and output channels.

Definition at line 128 of file base_convolution.hpp.

◆ m_dilations

template<typename TensorDataType , El::Device Device>
std::vector<int> lbann::base_convolution_layer< TensorDataType, Device >::m_dilations
protected

Convolution dilations.

Definition at line 134 of file base_convolution.hpp.

◆ m_groups

template<typename TensorDataType , El::Device Device>
int lbann::base_convolution_layer< TensorDataType, Device >::m_groups
protected

Convolution groups. The channels are split into this many independent groups when performing convolution. The default convolution operation has one group, and a depthwise convolution has as many groups as there are input channels.

Definition at line 140 of file base_convolution.hpp.

◆ m_output_channels

template<typename TensorDataType , El::Device Device>
int lbann::base_convolution_layer< TensorDataType, Device >::m_output_channels
protected

Definition at line 124 of file base_convolution.hpp.

◆ m_pads

template<typename TensorDataType , El::Device Device>
std::vector<int> lbann::base_convolution_layer< TensorDataType, Device >::m_pads
protected

Convolution padding.

Definition at line 130 of file base_convolution.hpp.

◆ m_strides

template<typename TensorDataType , El::Device Device>
std::vector<int> lbann::base_convolution_layer< TensorDataType, Device >::m_strides
protected

Convolution strides.

Definition at line 132 of file base_convolution.hpp.


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