cuda_mlp::CudaNetwork Class Reference

Feed-forward dense network with GPU-backed parameters and gradients. More...

Public Member Functions
	CudaNetwork (CublasHandle &handle)
	Construct a network tied to a cuBLAS handle. More...
void	addLayer (int in, int out, ActivationType act)
	Append a layer definition. More...
void	bindParams (unsigned int seed=kDefaultSeed)
	Allocate parameter/gradient buffers and initialize weights. More...
size_t	params_size () const
	Total number of parameters. More...
int	output_size () const
	Output dimension of the last layer. More...
CudaScalar *	params_data ()
	Mutable device pointer to parameters. More...
CudaScalar *	grads_data ()
	Mutable device pointer to gradients. More...
void	zeroGrads ()
	Zero all gradients. More...
void	forward_only (const CudaScalar *input, int batch)
	Forward pass only (no gradient computation) More...
CudaScalar	compute_loss_and_grad (const CudaScalar *input, const CudaScalar *target, int batch)
	Compute MSE loss and gradients for a batch. More...
void	copy_output_to_host (CudaScalar *host, size_t n) const
	Copy the latest output activations to host memory. More...
int	last_batch () const
	Batch size used in the last forward pass. More...

Detailed Description

Feed-forward dense network with GPU-backed parameters and gradients.

Constructor & Destructor Documentation

CudaNetwork()

cuda_mlp::CudaNetwork::CudaNetwork ( CublasHandle &  handle )

inlineexplicit

Construct a network tied to a cuBLAS handle.

Member Function Documentation

addLayer()

void cuda_mlp::CudaNetwork::addLayer ( int  in, int  out, ActivationType  act  )

inline

Append a layer definition.

Parameters

in	Input dimension
out	Output dimension
act	Activation function

bindParams()

void cuda_mlp::CudaNetwork::bindParams ( unsigned int  seed = kDefaultSeed )

inline

Allocate parameter/gradient buffers and initialize weights.

Parameters

seed	RNG seed for weight initialization

Here is the call graph for this function:

compute_loss_and_grad()

CudaScalar cuda_mlp::CudaNetwork::compute_loss_and_grad ( const CudaScalar *  input, const CudaScalar *  target, int  batch  )

inline

Compute MSE loss and gradients for a batch.

Parameters

input	Input batch (in x batch)
target	Target batch (out x batch)
batch	Batch size

Returns

Mean squared error loss

Here is the call graph for this function:

copy_output_to_host()

void cuda_mlp::CudaNetwork::copy_output_to_host ( CudaScalar *  host, size_t  n  ) const

inline

Copy the latest output activations to host memory.

forward_only()

void cuda_mlp::CudaNetwork::forward_only ( const CudaScalar *  input, int  batch  )

inline

Forward pass only (no gradient computation)

Parameters

input	Input batch (in x batch)
batch	Batch size

Here is the caller graph for this function:

grads_data()

CudaScalar* cuda_mlp::CudaNetwork::grads_data ( )

inline

Mutable device pointer to gradients.

Here is the call graph for this function:

last_batch()

int cuda_mlp::CudaNetwork::last_batch ( ) const

inline

Batch size used in the last forward pass.

output_size()

int cuda_mlp::CudaNetwork::output_size ( ) const

inline

Output dimension of the last layer.

Here is the caller graph for this function:

params_data()

CudaScalar* cuda_mlp::CudaNetwork::params_data ( )

inline

Mutable device pointer to parameters.

Here is the call graph for this function:

params_size()

size_t cuda_mlp::CudaNetwork::params_size ( ) const

inline

Total number of parameters.

zeroGrads()

void cuda_mlp::CudaNetwork::zeroGrads ( )

inline

Zero all gradients.

Here is the call graph for this function:

Here is the caller graph for this function:

---

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

• src/cuda/network.cuh