softmax

Documentation

Index

• type Layer
  • func StageAccuratePerChannel(options *OpMultiplier) *Layer
  • func StageAccuratePerInstance(options *OpMultiplier) *Layer
  • func StageFastPerChannel(options *OpMultiplier) *Layer
  • func StageFastPerInstance(options *OpMultiplier) *Layer
  • func StageLogPerChannel(options *OpMultiplier) *Layer
  • func StageLogPerInstance(options *OpMultiplier) *Layer
  • func (l *Layer) BackProp(handle *cudnn.Handler, dx, dy, y *layers.Tensor) error
  • func (l *Layer) ForwardProp(handle *cudnn.Handler, x, y *layers.Tensor) error
  • func (l *Layer) SetBackwardScalars(alpha, beta float64)
  • func (l *Layer) SetForwardScalars(alpha, beta float64)
• type OpMultiplier
• type Settings

Types

type Layer

type Layer struct {
	// contains filtered or unexported fields
}

Layer is a layer that holds the algos for softmax

func StageAccuratePerChannel

func StageAccuratePerChannel(options *OpMultiplier) *Layer

StageAccuratePerChannel stages the op to do a accurate softmax per channel. Options are alphas and betas forward and backward Forward prop will be: y= alpha * softmaxforward(x,y) + beta * y Backward will be: x= alpha * softmaxbackward(x,y) + beta * x Default values are:

Forward Prop : alpha = 1.0, beta = 0.0
Backward Prop: alpha = -1.0, beta = 0.0

func StageAccuratePerInstance

func StageAccuratePerInstance(options *OpMultiplier) *Layer

StageAccuratePerInstance stages the op to do a accurate softmax per instance. Options are alphas and betas forward and backward Forward prop will be: y= alpha * softmaxforward(x,y) + beta * y Backward will be: x= alpha * softmaxbackward(x,y) + beta * x Default values are:

Forward Prop : alpha = 1.0, beta = 0.0
Backward Prop: alpha = -1.0, beta = 0.0

func StageFastPerChannel

func StageFastPerChannel(options *OpMultiplier) *Layer

StageFastPerChannel stages the op to do a fast softmax per channel. Options are alphas and betas forward and backward Forward prop will be: y= alpha * softmaxforward(x,y) + beta * y Backward will be: x= alpha * softmaxbackward(x,y) + beta * x Default values are:

Forward Prop : alpha = 1.0, beta = 0.0
Backward Prop: alpha = -1.0, beta = 0.0

func StageFastPerInstance

func StageFastPerInstance(options *OpMultiplier) *Layer

StageFastPerInstance stages the op to do a fast softmax per instance. Options are alphas and betas forward and backward Forward prop will be: y= alpha * softmaxforward(x,y) + beta * y Backward will be: x= alpha * softmaxbackward(x,y) + beta * x Default values are:

Forward Prop : alpha = 1.0, beta = 0.0
Backward Prop: alpha = -1.0, beta = 0.0

func StageLogPerChannel

func StageLogPerChannel(options *OpMultiplier) *Layer

StageLogPerChannel stages the op to do a log softmax per channel. Options are alphas and betas forward and backward Forward prop will be: y= alpha * softmaxforward(x,y) + beta * y Backward will be: x= alpha * softmaxbackward(x,y) + beta * x Default values are:

Forward Prop : alpha = 1.0, beta = 0.0
Backward Prop: alpha = -1.0, beta = 0.0

func StageLogPerInstance

func StageLogPerInstance(options *OpMultiplier) *Layer

StageLogPerInstance stages the op to do a log softmax per intance. Options are alphas and betas forward and backward Forward prop will be: y= alpha * softmaxforward(x,y) + beta * y Backward will be: x= alpha * softmaxbackward(x,y) + beta * x Default values are:

Forward Prop : alpha = 1.0, beta = 0.0
Backward Prop: alpha = -1.0, beta = 0.0

func (*Layer) BackProp

func (l *Layer) BackProp(handle *cudnn.Handler, dx, dy, y *layers.Tensor) error

BackProp performs the backward propigation // x is the output

func (*Layer) ForwardProp

func (l *Layer) ForwardProp(handle *cudnn.Handler, x, y *layers.Tensor) error

ForwardProp performs the forward propigation y is the output

func (*Layer) SetBackwardScalars

func (l *Layer) SetBackwardScalars(alpha, beta float64)

SetBackwardScalars sets the backward alpha,beta scalars

func (*Layer) SetForwardScalars

func (l *Layer) SetForwardScalars(alpha, beta float64)

SetForwardScalars sets the forward alpha beta scalars

type OpMultiplier

type OpMultiplier struct {
	ForwardAlpha  float64
	ForwardBeta   float64
	BackwardAlpha float64
	BackwardBeta  float64
}

OpMultiplier sets the alpha beta scalars for the forward and backward operations

type Settings

type Settings struct {
	Mode gocudnn.SoftMaxMode      `json:"mode,omitempty"`
	Algo gocudnn.SoftMaxAlgorithm `json:"algo,omitempty"`
}

Settings contains all the info that is needed in order to perform the softmaxoutput