cnn

package
v0.0.0-...-98db5b7 Latest Latest
Warning

This package is not in the latest version of its module.

 Go to latest Published: May 7, 2020 License: CC-BY-4.0 Imports: 13 Imported by: 5

Documentation

Overview

Package cnn contains structs and methods used to do forward, and backward operations for convolution layers

Index

Constants

This section is empty.

Variables

This section is empty.

Functions

This section is empty.

Types

type Info 

type Info struct {
	Convolution convolution.Info `json:"convolution,omitempty"`
	Weights     tensor.Info      `json:"weights,omitempty"`
	Bias        tensor.Info      `json:"bias,omitempty"`
}

Info contains the info that is needed to build a CNN layer

type Layer 

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

Layer is a struct that holds filter, bias and convolution descriptors. The memory for w, dw, bias, dbias. The algos for forward, backward (data, filter) and the scalars for those algos. 1

func Setup 

func Setup(handle *cudnn.Handler,
	frmt gocudnn.TensorFormat,
	dtype gocudnn.DataType,
	mtype gocudnn.MathType,
	groupcount int32,
	filterdims []int32,
	convmode gocudnn.ConvolutionMode,
	pad,
	stride,
	dilation []int32,
	seed uint64) (*Layer, error)

Setup sets up the speed of the fwd and bwd algos dynamically. guessinputdims is really for setting up the random weights.

func SetupBasic 

func SetupBasic(handle *cudnn.Handler,
	frmt gocudnn.TensorFormat,
	dtype gocudnn.DataType,
	mtype gocudnn.MathType,
	groupcount int32,
	w, dw, b, db *layers.Tensor,
	convmode gocudnn.ConvolutionMode,
	pad,
	stride,
	dilation []int32) (*Layer, error)

SetupBasic sets up a convolution layer with the memory for the gpu added to it. This can be used for layers that share the same memory, but might have different convolution properties.

func (*Layer) BAvg 

func (c *Layer) BAvg(handle *cudnn.Handler) (float32, error)

BAvg returns the avg weight value for the layer

func (*Layer) BMax 

func (c *Layer) BMax(handle *cudnn.Handler) (float32, error)

BMax returns the Max bias value for the layer

func (*Layer) BMin 

func (c *Layer) BMin(handle *cudnn.Handler) (float32, error)

BMin returns the Min bias value for the layer

func (*Layer) BNorm1 

func (c *Layer) BNorm1(handle *cudnn.Handler) (float32, error)

BNorm1 returns the norm1 bias value for the layer

func (*Layer) BNorm2 

func (c *Layer) BNorm2(handle *cudnn.Handler) (float32, error)

BNorm2 returns the norm2 bias value for the layer

func (*Layer) BackPropData 

func (c *Layer) BackPropData(handle *cudnn.Handler, wspace *nvidia.Malloced, dx, dy *layers.Tensor) error

BackPropData performs the BackPropData

func (*Layer) BackPropFilter 

func (c *Layer) BackPropFilter(handle *cudnn.Handler, wspace *nvidia.Malloced, x, dy *layers.Tensor) error

BackPropFilter does the backward propagation for the filter You will pass a handle workspace memory x,dy layer.io

func (*Layer) BackPropFilterData 

func (c *Layer) BackPropFilterData(handle *cudnn.Handler, wspacedata, wspacefilter *nvidia.Malloced, x, dx, dy *layers.Tensor) error

BackPropFilterData does the backprop for the data and the filter

Dx and X can be the same.

func (*Layer) Bias 

func (c *Layer) Bias() *layers.Tensor

Bias returns the Bias

func (*Layer) DBAvg 

func (c *Layer) DBAvg(handle *cudnn.Handler) (float32, error)

DBAvg returns the avg delta bias value for the layer

func (*Layer) DBMax 

func (c *Layer) DBMax(handle *cudnn.Handler) (float32, error)

DBMax returns the Max delta bias value for the layer

func (*Layer) DBMin 

func (c *Layer) DBMin(handle *cudnn.Handler) (float32, error)

DBMin returns the Min delta bias value for the layer

func (*Layer) DBNorm1 

func (c *Layer) DBNorm1(handle *cudnn.Handler) (float32, error)

DBNorm1 returns the norm1 delta bias value for the layer

func (*Layer) DBNorm2 

func (c *Layer) DBNorm2(handle *cudnn.Handler) (float32, error)

DBNorm2 returns the norm2 delta bias value for the layer

func (*Layer) DWAvg 

func (c *Layer) DWAvg(handle *cudnn.Handler) (float32, error)

DWAvg returns the avg delta weight value for the layer

func (*Layer) DWMax 

func (c *Layer) DWMax(handle *cudnn.Handler) (float32, error)

DWMax returns the Max delta weight value for the layer

func (*Layer) DWMin 

func (c *Layer) DWMin(handle *cudnn.Handler) (float32, error)

DWMin returns the Min delta weight value for the layer

func (*Layer) DWNorm1 

func (c *Layer) DWNorm1(handle *cudnn.Handler) (float32, error)

DWNorm1 returns the norm1 delta weight value for the layer

func (*Layer) DWNorm2 

func (c *Layer) DWNorm2(handle *cudnn.Handler) (float32, error)

DWNorm2 returns the norm2 delta weight value for the layer

func (*Layer) DeltaWeights 

func (c *Layer) DeltaWeights() *layers.Tensor

DeltaWeights returns the deltaweights

func (*Layer) DeltaWeightsFillSlice 

func (c *Layer) DeltaWeightsFillSlice(h *cudnn.Handler, input interface{}, length int) error

DeltaWeightsFillSlice will fill the weights with values

func (*Layer) FilterProps 

func (c *Layer) FilterProps() (gocudnn.TensorFormat, gocudnn.DataType, []int32, error)

FilterProps returns the filter properties of the Convolution Layer

func (*Layer) FindOutputDims 

func (c *Layer) FindOutputDims(input *layers.Tensor) (dims []int32, err error)

FindOutputDims finds the outputdims fo the cnn

func (*Layer) ForwardProp 

func (c *Layer) ForwardProp(handle *cudnn.Handler, wspace *nvidia.Malloced, x, y *layers.Tensor) error

ForwardProp performs the ForwardProp

func (*Layer) GetBwdDataAlgoPerfList 

func (c *Layer) GetBwdDataAlgoPerfList(handle *cudnn.Handler, x, y *layers.Tensor, workspace *nvidia.Malloced) ([]convolution.BackDataPerformance, error)

GetBwdDataAlgoPerfList gets a list of backward performance algos

func (*Layer) GetBwdFiltAlgoPerfList 

func (c *Layer) GetBwdFiltAlgoPerfList(handle *cudnn.Handler, x, y *layers.Tensor, workspace *nvidia.Malloced) ([]convolution.BackFilterPerformance, error)

GetBwdFiltAlgoPerfList gets a list of forward performance algos

func (*Layer) GetFwdAlgoPerfList 

func (c *Layer) GetFwdAlgoPerfList(handle *cudnn.Handler, x, y *layers.Tensor, workspace *nvidia.Malloced) ([]convolution.ForwardPerformance, error)

GetFwdAlgoPerfList gets a list of forward performance algos

func (*Layer) Info 

func (l *Layer) Info() (Info, error)

Info returns the info struct for the convolution

func (*Layer) L1L2Loss 

func (c *Layer) L1L2Loss() (L1 float32, L2 float32)

L1L2Loss will return the L1 loss and L2 loss for the layer

func (*Layer) LoadBiasValues 

func (c *Layer) LoadBiasValues(handle *cudnn.Handler, slice interface{}, length int) error

LoadBiasValues will load a slice into cuda memory for the Weights.

func (*Layer) LoadBiasValuesEX 

func (c *Layer) LoadBiasValuesEX(handle *cudnn.Handler, r io.Reader) error

LoadBiasValuesEX takes a reader and coppies the bytes over to the bias

func (*Layer) LoadTrainer 

func (c *Layer) LoadTrainer(handle *cudnn.Handler, forweights, forbias trainer.Trainer) error

LoadTrainer sets up the momentum trainer

func (*Layer) LoadWValues 

func (c *Layer) LoadWValues(handle *cudnn.Handler, slice interface{}, length int) error

LoadWValues will load a slice into cuda memory for the Weights.

func (*Layer) LoadWvaluesEX 

func (c *Layer) LoadWvaluesEX(handle *cudnn.Handler, r io.Reader) error

LoadWvaluesEX takes a reader and coppies the bytes over to the weights

func (*Layer) MakeOutputTensor 

func (c *Layer) MakeOutputTensor(handle *cudnn.Handler, input *layers.Tensor) (*layers.Tensor, error)

MakeOutputTensor makes the output tensor of the layer

func (*Layer) MakeOutputTensorInference 

func (c *Layer) MakeOutputTensorInference(handle *cudnn.Handler, input *layers.Tensor) (*layers.Tensor, error)

MakeOutputTensorInference makes the output tensor of the layer

func (*Layer) MakeRandom 

func (c *Layer) MakeRandom(h *cudnn.Handler, inputdims []int32) error

MakeRandom does what it says it will make the weights random considering the fanin

func (*Layer) OutputDims 

func (c *Layer) OutputDims(inputdims []int32) []int32

OutputDims will return the dims for the output

func (*Layer) SetBackwardScalars 

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

SetBackwardScalars sets the alpha and beta scalars, the defaults are alpha, 1, beta=0 and are initialized in the function FilterSetup

func (*Layer) SetBestAlgosConsidering 

func (c *Layer) SetBestAlgosConsidering(handle *cudnn.Handler, x, y *layers.Tensor, wspacelimit int, fastest bool) (uint, error)

SetBestAlgosConsidering this method will set the best algos for the fwd, bwddata, and bwdfilter algos. and return the workspace size along with an error if an error is found the function will not set any values, Here are some simple rules to the function if fastest is marked true. Then it will find the fastest algo no mater what worksize is. if fastest is set to false. It will check if wspace is greater than zero then it will set the algos to the fastest algo considering the workspace size, and return the largest wspacesize in all the algos else it will find and set the fastest algos with no workspace size and return 0

func (*Layer) SetBestAlgosConsideringDims4d 

func (c *Layer) SetBestAlgosConsideringDims4d(handle *cudnn.Handler, x, y, w []int32, wspacelimit int, fastest bool) (uint, error)

SetBestAlgosConsideringDims4d this method will set the best algos for the fwd, bwddata, and bwdfilter algos. and return the workspace size along with an error if an error is found the function will not set any values, Here are some simple rules to the function if fastest is marked true. Then it will find the fastest algo no mater what worksize is. if fastest is set to false. It will check if wspace is greater than zero then it will set the algos to the fastest algo considering the workspace size, and return the largest wspacesize in all the algos else it will find and set the fastest algos with no workspace size and return 0

func (*Layer) SetBwdDataAlgoPerformance 

func (c *Layer) SetBwdDataAlgoPerformance(bwddata convolution.BackDataPerformance)

SetBwdDataAlgoPerformance sets the Performance Values

func (*Layer) SetBwdFiltAlgoPerformance 

func (c *Layer) SetBwdFiltAlgoPerformance(bwdfilt convolution.BackFilterPerformance)

SetBwdFiltAlgoPerformance sets the Performance Values

func (*Layer) SetForwardScalars 

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

SetForwardScalars sets the alpha and beta scalars, the defaults are alpha, 1, beta=0 and are initialized in the function FilterSetup

func (*Layer) SetFwdAlgoPerformance 

func (c *Layer) SetFwdAlgoPerformance(fwd convolution.ForwardPerformance)

SetFwdAlgoPerformance sets the Performance Values

func (*Layer) SetMathType 

func (c *Layer) SetMathType(mtype gocudnn.MathType) error

SetMathType sets the mathtype

func (*Layer) SetOtherScalars 

func (c *Layer) SetOtherScalars(alpha, beta float64)

SetOtherScalars sets the alpha and beta scalars for the weights, the defaults are alpha, 1, beta=1 and are initialized in the function FilterSetup

func (*Layer) String 

func (c *Layer) String() string

func (*Layer) ToggleBiasPrintValueForStringer 

func (c *Layer) ToggleBiasPrintValueForStringer()

ToggleBiasPrintValueForStringer toggles if Bias values will be printed

func (*Layer) ToggleDBiasPrintValueForStringer 

func (c *Layer) ToggleDBiasPrintValueForStringer()

ToggleDBiasPrintValueForStringer toggles if dBias values will be printed

func (*Layer) ToggleDWeightsPrintValueForStringer 

func (c *Layer) ToggleDWeightsPrintValueForStringer()

ToggleDWeightsPrintValueForStringer toggles if DWeight values will be printed

func (*Layer) ToggleWeightsPrintValueForStringer 

func (c *Layer) ToggleWeightsPrintValueForStringer()

ToggleWeightsPrintValueForStringer toggles if Weight values will be printed

func (*Layer) UpdateWeights 

func (c *Layer) UpdateWeights(handle *cudnn.Handler, batch, epoch int) error

UpdateWeights does the weight update

func (*Layer) WAvg 

func (c *Layer) WAvg(handle *cudnn.Handler) (float32, error)

WAvg returns the avg weight value for the layer

func (*Layer) WMax 

func (c *Layer) WMax(handle *cudnn.Handler) (float32, error)

WMax returns the Max weight value for the layer.

func (*Layer) WMin 

func (c *Layer) WMin(handle *cudnn.Handler) (float32, error)

WMin returns the Min weight value for the layer

func (*Layer) WNorm1 

func (c *Layer) WNorm1(handle *cudnn.Handler) (float32, error)

WNorm1 returns the norm1 weight value for the layer

func (*Layer) WNorm2 

func (c *Layer) WNorm2(handle *cudnn.Handler) (float32, error)

WNorm2 returns the norm2 weight value for the layer

func (*Layer) Weights 

func (c *Layer) Weights() *layers.Tensor

Weights returns the weights

func (*Layer) WeightsFillSlice 

func (c *Layer) WeightsFillSlice(h *cudnn.Handler, input interface{}, length int) error

WeightsFillSlice will fill a slice with the weight values

Source Files

View all Source files

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL
go.dev uses cookies from Google to deliver and enhance the quality of its services and to analyze traffic. Learn more.