cuda_mlp::CudaMinimizerBase Class Reference

Abstract base class for CUDA-based minimizers. More...

Inheritance diagram for cuda_mlp::CudaMinimizerBase:

Collaboration diagram for cuda_mlp::CudaMinimizerBase:

Public Types
using	LossGradFun = std::function< CudaScalar(const CudaScalar *params, CudaScalar *grad, const CudaScalar *input, const CudaScalar *target, int batch)>
	Loss and gradient callback signature. More...
using	IterHook = std::function< void(int)>
	Optional per-iteration hook signature. More...

Public Member Functions
	CudaMinimizerBase (CublasHandle &handle)
	Construct with a cuBLAS handle reference. More...
virtual	~CudaMinimizerBase ()=default
int	iterations () const noexcept
	Return the number of iterations performed in the last solve. More...
void	setRecorder (::IterationRecorder< CudaBackend > *recorder)
	Attach a recorder for loss/grad norm history. More...
void	setMaxIterations (int iters)
	Set maximum number of iterations. More...
void	setTolerance (CudaScalar tol)
	Set stopping tolerance (interpretation depends on optimizer) More...
void	setLineSearchParams (int max_iters, CudaScalar c1, CudaScalar rho)
	Configure Armijo line search parameters. More...
virtual void	solve (int n, CudaScalar *params, const CudaScalar *input, const CudaScalar *target, int batch, const LossGradFun &loss_grad)=0
	Solve the optimization problem. More...

Protected Attributes
CublasHandle &	handle_
	cuBLAS handle used by the optimizer More...
int	max_iters_ = 200
int	max_line_iters_ = 20
	Iteration limits. More...
CudaScalar	tol_ = 1e-6f
CudaScalar	c1_ = 1e-4f
CudaScalar	rho_ = 0.5f
	Stopping and line-search params. More...
int	last_iterations_ = 0
	Iterations performed in last run. More...
::IterationRecorder< CudaBackend > *	recorder_ = nullptr
	Optional recorder for diagnostics. More...

Detailed Description

Abstract base class for CUDA-based minimizers.

Member Typedef Documentation

IterHook

using cuda_mlp::CudaMinimizerBase::IterHook = std::function<void(int)>

Optional per-iteration hook signature.

LossGradFun

using cuda_mlp::CudaMinimizerBase::LossGradFun = std::function<CudaScalar( const CudaScalar *params, CudaScalar *grad, const CudaScalar *input, const CudaScalar *target, int batch)>

Loss and gradient callback signature.

Constructor & Destructor Documentation

CudaMinimizerBase()

cuda_mlp::CudaMinimizerBase::CudaMinimizerBase ( CublasHandle &  handle )

inlineexplicit

Construct with a cuBLAS handle reference.

~CudaMinimizerBase()

virtual cuda_mlp::CudaMinimizerBase::~CudaMinimizerBase ( )

virtualdefault

Member Function Documentation

iterations()

int cuda_mlp::CudaMinimizerBase::iterations ( ) const

inlinenoexcept

Return the number of iterations performed in the last solve.

setLineSearchParams()

void cuda_mlp::CudaMinimizerBase::setLineSearchParams ( int  max_iters, CudaScalar  c1, CudaScalar  rho  )

inline

Configure Armijo line search parameters.

Parameters

max_iters	Maximum line-search iterations
c1	Armijo sufficient decrease constant
rho	Backtracking factor in (0,1)

setMaxIterations()

void cuda_mlp::CudaMinimizerBase::setMaxIterations ( int  iters )

inline

Set maximum number of iterations.

setRecorder()

void cuda_mlp::CudaMinimizerBase::setRecorder ( ::IterationRecorder< CudaBackend > *  recorder )

inline

Attach a recorder for loss/grad norm history.

setTolerance()

void cuda_mlp::CudaMinimizerBase::setTolerance ( CudaScalar  tol )

inline

Set stopping tolerance (interpretation depends on optimizer)

solve()

virtual void cuda_mlp::CudaMinimizerBase::solve ( int  n, CudaScalar *  params, const CudaScalar *  input, const CudaScalar *  target, int  batch, const LossGradFun &  loss_grad  )

pure virtual

Solve the optimization problem.

Parameters

n	Number of parameters
params	Parameter vector (device pointer)
input	Input data (device pointer)
target	Target data (device pointer)
batch	Batch size
loss_grad	Callback that returns loss and writes gradient

Implemented in cuda_mlp::CudaSGD, cuda_mlp::CudaGD, and cuda_mlp::CudaLBFGS.

Member Data Documentation

c1_

CudaScalar cuda_mlp::CudaMinimizerBase::c1_ = 1e-4f

protected

handle_

CublasHandle& cuda_mlp::CudaMinimizerBase::handle_

protected

cuBLAS handle used by the optimizer

last_iterations_

int cuda_mlp::CudaMinimizerBase::last_iterations_ = 0

protected

Iterations performed in last run.

max_iters_

int cuda_mlp::CudaMinimizerBase::max_iters_ = 200

protected

max_line_iters_

int cuda_mlp::CudaMinimizerBase::max_line_iters_ = 20

protected

Iteration limits.

recorder_

::IterationRecorder<CudaBackend>* cuda_mlp::CudaMinimizerBase::recorder_ = nullptr

protected

Optional recorder for diagnostics.

rho_

CudaScalar cuda_mlp::CudaMinimizerBase::rho_ = 0.5f

protected

Stopping and line-search params.

tol_

CudaScalar cuda_mlp::CudaMinimizerBase::tol_ = 1e-6f

protected

---

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

• src/cuda/minimizer_base.cuh