LocalUpdater do gradient aggregation and update gradient calling the wrapped Optimizer on a specific device (i.e., CPU or GPU). More...

#include <updater.h>

Inheritance diagram for singa::LocalUpdater:

Collaboration diagram for singa::LocalUpdater:

Public Member Functions
	LocalUpdater (int total_num, Optimizer *opt, std::shared_ptr< Device > dev=defaultDevice)

virtual void	Register (const string &name, const ParamSpec &specs) override
	Forward Register() to Optimizer.

virtual void	Apply (int step, const string &name, Tensor &grad, Tensor &value) override
	Update parameter value based on given gradient by invoking optimizer algoritim. More...

Public Member Functions inherited from singa::Updater
	Updater (Optimizer *opt)

virtual void	Setup (const OptimizerConf &conf)
	Forward Setup() to Optimizer.

Optimizer *	GetOptimizer ()

	Updater (const Updater &)=delete

void	operator= (const Updater &)=delete

Additional Inherited Members
Protected Attributes inherited from singa::Updater
Optimizer *	opt_

Detailed Description

LocalUpdater do gradient aggregation and update gradient calling the wrapped Optimizer on a specific device (i.e., CPU or GPU).

Member Function Documentation

virtual void singa::LocalUpdater::Apply	(	int	step,
		const string &	name,
		Tensor &	grad,
		Tensor &	value
	)

overridevirtual

Update parameter value based on given gradient by invoking optimizer algoritim.

When tranining net call this function will be blocked until all the partial gradients are aggrageted in a synchronized style training.

Reimplemented from singa::Updater.

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