Package deepnetts.cudnn

Class FullyConnectedBackwardCuda

java.lang.Object

deepnetts.cudnn.AbstractBackwardCuda

deepnetts.cudnn.FullyConnectedBackwardCuda

All Implemented Interfaces:

BackwardAccelerator

public final class FullyConnectedBackwardCuda
extends AbstractBackwardCuda

Constructor Summary

Constructors

Constructor	Description
FullyConnectedBackwardCuda(CudaHandles cudaHandles, FullyConnectedLayer layer)
FullyConnectedBackwardCuda(CudaHandles cudaHandles, TensorBase outputs, TensorBase weights, TensorBase biases, TensorBase inputs)

Method Summary

Modifier and Type	Method	Description
void	backward()	Performs forward pass.

Methods inherited from class deepnetts.cudnn.AbstractBackwardCuda

copyTensorsToDevice

Methods inherited from class java.lang.Object

equals, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Details

FullyConnectedBackwardCuda

public FullyConnectedBackwardCuda(CudaHandles cudaHandles,
 TensorBase outputs,
 TensorBase weights,
 TensorBase biases,
 TensorBase inputs)

Parameters:

cudaHandles - cuda library context handles

outputs - 1D tensor with outputs for this layer

weights - 2D tensor with weights

biases - 1D tensor with biases

inputs - 1D tensor with inputs from previous layer maybe change order of tensors? i, w, o | mada out = W * X where * is dot product make sure that outputs contains biases and the beta is 1 Ovo izbaci imas novi konstruktor ispod koji je zamena ikoji mora da ima referncu na layer KOristim je samo za testiranje

FullyConnectedBackwardCuda

public FullyConnectedBackwardCuda(CudaHandles cudaHandles,
 FullyConnectedLayer layer)

Method Details

backward

public void backward()

Description copied from class: AbstractBackwardCuda

Performs forward pass.

Specified by:

backward in interface BackwardAccelerator

Specified by:

backward in class AbstractBackwardCuda