RecurrentNetworkExecutor is a specialized runtime for recurrent neural networks (RNNs). More...

#include <recurrent_network_executor.h>

Inheritance diagram for caffe2::RecurrentNetworkExecutorBase:

Public Member Functions
virtual bool	Run (int T)=0

virtual bool	RunBackwards (int T)=0

void	EnsureTimestepInitialized (int t, Workspace *ws, const std::vector< std::unique_ptr< ObserverBase< OperatorBase >>> &observers_list)
	Callers must call EnsureTimestepInitialized before starting execution for each of the relevant timesteps. More...

void	SetMaxParallelTimesteps (int p)
	Set limit for the number of timesteps that run in parallel. More...

size_t	NumObserversStepNet ()

Data Fields
bool	debug_ = false

Protected Member Functions
	RecurrentNetworkExecutorBase (const NetDef &step_net_def, std::map< string, string > &recurrent_input_map, std::string timestep_blob)

void	PrintInfo (int t)
	For debug purposes, print the dependency structure. More...

virtual void	AnalyzeOps ()

virtual bool	ignoreLinkDependencies ()=0

Protected Attributes
std::vector< std::vector< RNNNetOperator > >	timestep_ops_

std::vector< OperatorBase * >	op_ptrs_

std::vector< RNNNetOperator >	timestep_ops_template_

NetDef	step_net_def_

std::vector< std::vector< string > >	op_deps_

std::vector< Workspace * >	workspaces_

std::map< string, string >	recurrent_input_map_

std::string	timestep_blob_

int	max_parallel_timesteps_ = -1

Detailed Description

RecurrentNetworkExecutor is a specialized runtime for recurrent neural networks (RNNs).

It is invoked from the RecurrentNetworkOp and RecurrentNetworkGradientOp.

Its main benefit over running each RNN timestep as a separate net is that it can run ops in subsequent timesteps in parallel when possible. For example, multi-layer LSTMs allow for timestep parallelism because next timestep's lower layer can start executing at the same time as the same timestep's upper layer.

There are two implementations of the RNN executor: one for CPUs (ThreadedRecurrentNetworkExecutor) and another for GPUs (CUDARecurrentNetworkExecutor).

Definition at line 31 of file recurrent_network_executor.h.

Member Function Documentation

void caffe2::RecurrentNetworkExecutorBase::EnsureTimestepInitialized	(	int	t,
		Workspace *	ws,
		const std::vector< std::unique_ptr< ObserverBase< OperatorBase >>> &	observers_list
	)

inline

Callers must call EnsureTimestepInitialized before starting execution for each of the relevant timesteps.

If timestep was initialized before, this is a no-op. First time this is called the dependencies of the operators in timestep are analyzed, and that incurs higher overhead than subsequent calls.

Definition at line 74 of file recurrent_network_executor.h.

void caffe2::RecurrentNetworkExecutorBase::PrintInfo ( int t )

inlineprotected

For debug purposes, print the dependency structure.

Set rnn_executor_debug=1 in the RecurrentNetworkOp to enable.

Definition at line 418 of file recurrent_network_executor.h.

void caffe2::RecurrentNetworkExecutorBase::SetMaxParallelTimesteps ( int p )

inline

Set limit for the number of timesteps that run in parallel.

Useful for forward-only execution when we rotate workspaces over timesteps, i.e when timestep[t] and timestep[t + p] have same workspace.

Definition at line 185 of file recurrent_network_executor.h.

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