cuda_mlp::CudaDenseLayer Class Reference

Fully connected layer with activation, using column-major matrices. More...

Public Member Functions
	CudaDenseLayer (int in, int out, ActivationType act)
	Construct a dense layer. More...
int	in () const
	Input dimension. More...
int	out () const
	Output dimension. More...
size_t	params_size () const
	Total parameter count (weights + bias) More...
size_t	weights_size () const
	Weights parameter count. More...
size_t	bias_size () const
	Bias parameter count. More...
void	bind (CudaScalar *params, CudaScalar *grads)
	Bind parameter and gradient buffers. More...
CudaScalar	init_stddev () const
	Recommended stddev for weight initialization. More...
void	forward (CublasHandle &handle, const CudaScalar *input, int batch, CudaScalar *output)
	Forward pass: Z = W*X + b, A = act(Z) More...
void	backward (CublasHandle &handle, const CudaScalar *input, const CudaScalar *output, CudaScalar *next_grad, int batch, CudaScalar *prev_grad)
	Backward pass: compute dW, db, and optionally dX. More...
const CudaScalar *	params_ptr () const
	Raw parameter pointer for this layer. More...
const CudaScalar *	grads_ptr () const
	Raw gradient pointer for this layer. More...

Detailed Description

Fully connected layer with activation, using column-major matrices.

Constructor & Destructor Documentation

CudaDenseLayer()

cuda_mlp::CudaDenseLayer::CudaDenseLayer ( int  in, int  out, ActivationType  act  )

inline

Construct a dense layer.

Parameters

in	Input dimension
out	Output dimension
act	Activation type

Member Function Documentation

backward()

void cuda_mlp::CudaDenseLayer::backward ( CublasHandle &  handle, const CudaScalar *  input, const CudaScalar *  output, CudaScalar *  next_grad, int  batch, CudaScalar *  prev_grad  )

inline

Backward pass: compute dW, db, and optionally dX.

Parameters

handle	cuBLAS handle
input	Input activations
output	Output activations
next_grad	Gradient w.r.t. output (out x batch), updated in-place
batch	Batch size
prev_grad	Optional gradient w.r.t. input (in x batch)

Here is the call graph for this function:

bias_size()

size_t cuda_mlp::CudaDenseLayer::bias_size ( ) const

inline

Bias parameter count.

bind()

void cuda_mlp::CudaDenseLayer::bind ( CudaScalar *  params, CudaScalar *  grads  )

inline

Bind parameter and gradient buffers.

forward()

void cuda_mlp::CudaDenseLayer::forward ( CublasHandle &  handle, const CudaScalar *  input, int  batch, CudaScalar *  output  )

inline

Forward pass: Z = W*X + b, A = act(Z)

Parameters

handle	cuBLAS handle
input	Input matrix (in x batch)
batch	Batch size
output	Output matrix (out x batch)

Here is the call graph for this function:

grads_ptr()

const CudaScalar* cuda_mlp::CudaDenseLayer::grads_ptr ( ) const

inline

Raw gradient pointer for this layer.

in()

int cuda_mlp::CudaDenseLayer::in ( ) const

inline

Input dimension.

init_stddev()

CudaScalar cuda_mlp::CudaDenseLayer::init_stddev ( ) const

inline

Recommended stddev for weight initialization.

Here is the call graph for this function:

out()

int cuda_mlp::CudaDenseLayer::out ( ) const

inline

Output dimension.

params_ptr()

const CudaScalar* cuda_mlp::CudaDenseLayer::params_ptr ( ) const

inline

Raw parameter pointer for this layer.

params_size()

size_t cuda_mlp::CudaDenseLayer::params_size ( ) const

inline

Total parameter count (weights + bias)

weights_size()

size_t cuda_mlp::CudaDenseLayer::weights_size ( ) const

inline

Weights parameter count.

---

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

• src/cuda/layer.cuh