gocunets

package module
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: 24 Imported by: 4

README

gocunets

The kame ha me ha of neural networks in go.

Pull requests are welcomed.

packages needed

go get github.com/nfnt/resize     // Will eventually get rid of this.
go get github.com/dereklstinson/gpu-monitoring-tools/bindings/go/nvml  //Will eventually get bindings in gocudnn. 
go get github.com/pkg/browser     // This is used with ui it auto launches browser.  Not so useful when used with a headless machine
go get github.com/dereklstinson/nccl
go get github.com/dereklstinson/gocudnn
go get github.com/dereklstinson/cutil
go get -u gonum.org/v1/gonum/...  // Will eventually get rid of this
go get gonum.org/v1/plot/...      // Will want to get rid of this and use javascript plotting.

GoCuNets is basically 100% GPU computing using the gocudnn package. Right now the only gpu support is Nvidia. Eventually, AMD GPUs will have support through HIP and MIOpen. I want to make it so that when using this package you don't have to download both cuda and hip libraries.

This package is separated into a few parts parts

github.com/dereklstinson/gocunets/devices
github.com/dereklstinson/gocunets/layers
github.com/dereklstinson/gocunets/loss
github.com/dereklstinson/gocunets/ui
github.com/dereklstinson/gocunets/trainer
github.com/dereklstinson/gocunets

The sub-package devices contains sub-packages for wrappers to device library bindings.

The sub-package layers contains other sub-packages to layers.

The sub-package loss contains different loss algorithms.

The sub-package ui contains a makeshift real time neural network monitoring server. Eventually, I would like to have it create a report / reports.

The sub-package trainer contains weight trainers.

The main package contains a higher level interface.

More on GoCuNets

A lot has changed since the beginning of GoCuNets. A lot of features that used to be available are not available for the moment. Like ui interface and model saving and loading. Saving and loading will be implemented soon. It will be part an upcoming Model interface.

type Builder has exposed flags that can be set through their methods. You only need to set the ones that are going to be used. A builder builds layers, tensors, randomly initialized tensors and more. This makes it easier to build networks without having pass flags all the time.

These flags are

    Frmt TensorFormat
	Dtype     DataType
	Cmode     ConvolutionMode
	Mtype     MathType
	Pmode     PoolingMode
	AMode     ActivationMode
	BNMode    BatchNormMode
	Nan       NanProp
```text

Type Handle is a handle to a GPU.  A GPU can have more than one Handle.  One worker can be used on multiple handles.
Handle and worker must use the same device.

Module interface is new.  Modules will eventually be used to implement graphs.  There are a few modules that I have made.
Eventually everything from layer sub-package will have wrappers in gocunets and will be made into a module.  

This is a working package, but it is pre alpha.
Better version management will be coming.

Documentation

Index

Constants

This section is empty.

Variables

View Source 
var Flags struct {
	Format TensorFormat
	Dtype  DataType
	Nan    NanProp
	CMode  ConvolutionMode
	BNMode BatchNormMode
	BNOps  BatchNormOps
	PMode  PoolingMode
	AMode  ActivationMode
	SMMode SoftmaxMode
	SMAlgo SoftmaxAlgo
	MType  MathType
}

Flags is a struct that should only be used for passing flags.

Functions

func BackPropForSharedInputForModuleNetworks 

func BackPropForSharedInputForModuleNetworks(m []*SimpleModuleNetwork) (err error)

BackPropForSharedInputForModuleNetworks is a hack to make up if two module networks share the same input. It will zero out the dx values for the module and then run back propagation

func ModuleActivationDebug 

func ModuleActivationDebug()

ModuleActivationDebug is for debugging

func ModuleBackwardDataDebug 

func ModuleBackwardDataDebug()

ModuleBackwardDataDebug is for debugging

func ModuleBackwardFilterDebug 

func ModuleBackwardFilterDebug()

ModuleBackwardFilterDebug is for debugging

func ModuleConcatDebug 

func ModuleConcatDebug()

ModuleConcatDebug is for debugging

func ModuleForwardDebug 

func ModuleForwardDebug()

ModuleForwardDebug is for debugging

func PerformanceDebugging 

func PerformanceDebugging()

PerformanceDebugging - if function called it raises flag to print performance for inner layers

func SetPeerAccess 

func SetPeerAccess(devs []Device) (connections int, err error)

SetPeerAccess sets peer access accross all devices

Types

type ActMode 

type ActMode int

ActMode are the activationmode flags

type ActivaitonModeFlag 

type ActivaitonModeFlag struct {
}

ActivaitonModeFlag passes ActivationMode flags

func (ActivaitonModeFlag) Logistic 

func (a ActivaitonModeFlag) Logistic() ActMode

Logistic returns ActivationMode flag for Logistic

func (ActivaitonModeFlag) SoftMax 

func (a ActivaitonModeFlag) SoftMax() ActMode

SoftMax returns ActivationMode flag for softmax

func (ActivaitonModeFlag) Tanh 

func (a ActivaitonModeFlag) Tanh() ActMode

Tanh returns ActivationMode flag for Tanh

type ActivationMode 

type ActivationMode struct {
	act.Mode
}

ActivationMode struct wrapper for gocudnn.ActivationMode. Look up methods in gocudnn.

func (*ActivationMode) ClippedRelu 

func (a *ActivationMode) ClippedRelu() ActivationMode

ClippedRelu sets and returns the ClippedRelu flag

func (*ActivationMode) Elu 

func (a *ActivationMode) Elu() ActivationMode

Elu sets and returns the Elu flag

func (*ActivationMode) Identity 

func (a *ActivationMode) Identity() ActivationMode

Identity sets and returns the Identity flag

func (*ActivationMode) Leaky 

func (a *ActivationMode) Leaky() ActivationMode

Leaky sets and returns the Leaky flag

func (*ActivationMode) PRelu 

func (a *ActivationMode) PRelu() ActivationMode

PRelu sets and returns the PRelu flag

func (*ActivationMode) Relu 

func (a *ActivationMode) Relu() ActivationMode

Relu sets and returns the Relu flag

func (*ActivationMode) Sigmoid 

func (a *ActivationMode) Sigmoid() ActivationMode

Sigmoid sets and returns the Sigmoid flag

func (*ActivationMode) Tanh 

func (a *ActivationMode) Tanh() ActivationMode

Tanh sets and returns the Tanh flag

func (*ActivationMode) Threshhold 

func (a *ActivationMode) Threshhold() ActivationMode

Threshhold sets and returns the Threshhold flag

type Backwarder 

type Backwarder interface {
	Backward() error
}

Backwarder does the backward operation

type BatchNormMode 

type BatchNormMode struct {
	gocudnn.BatchNormMode
}

BatchNormMode struct wrapper for gocudnn.BatchNormMode. Look up methods in gocudnn.

func (*BatchNormMode) PerActivation 

func (b *BatchNormMode) PerActivation() BatchNormMode

PerActivation sets and returns the PerActivation flag

func (*BatchNormMode) Spatial 

func (b *BatchNormMode) Spatial() BatchNormMode

Spatial sets and returns the Spatial flag

func (*BatchNormMode) SpatialPersistent 

func (b *BatchNormMode) SpatialPersistent() BatchNormMode

SpatialPersistent sets and returns the SpatialPersistent flag

type BatchNormOps 

type BatchNormOps struct {
	gocudnn.BatchNormOps
}

BatchNormOps struct wrapper for gocudnn.BatchNormOps. Look up methods in gocudnn.

func (*BatchNormOps) Activation 

func (b *BatchNormOps) Activation() BatchNormOps

Activation sets and returns the Activation flag

func (*BatchNormOps) AddActivation 

func (b *BatchNormOps) AddActivation() BatchNormOps

AddActivation sets and returns the AddActivation flag

func (*BatchNormOps) Normal 

func (b *BatchNormOps) Normal() BatchNormOps

Normal sets and returns the Normal flag

type Builder 

type Builder struct {
	Frmt   TensorFormat
	Dtype  DataType
	Cmode  ConvolutionMode
	Mtype  MathType
	Pmode  PoolingMode
	AMode  ActivationMode
	BNMode BatchNormMode
	Nan    NanProp
	// contains filtered or unexported fields
}

Builder will create layers with the flags set within the struct

func CreateBuilder 

func CreateBuilder(h *Handle) (b *Builder)

CreateBuilder creates a Builder. Flags can be set by flags' methods inside of Builder. Default Flags are set at: 

	Frmt.NCHW()

	Mtype.Default()

	Nan.NotPropigate()

	Cmode.CrossCorrelation()

	Dtype.Float()

	Pmode.AverageCountExcludePadding()

 BNMode.Spatial()

	AMode.Leaky()

func (*Builder) Activation 

func (l *Builder) Activation(id int64) (a *Layer, err error)

Activation creates an activation layer

func (*Builder) AllocateMemory 

func (l *Builder) AllocateMemory(sib uint) (cutil.Pointer, error)

AllocateMemory allocates memory

func (*Builder) BatchNorm 

func (l *Builder) BatchNorm(id int64) (batch *Layer, err error)

BatchNorm is the batch norm layer

func (*Builder) ConnectLayers 

func (l *Builder) ConnectLayers(layer1, layer2 *Layer) error

ConnectLayers Creates the output of layer1 and connects it as the input to layer2.

func (*Builder) ConvolutionLayer 

func (l *Builder) ConvolutionLayer(id int64, groupcount int32, w, dw, b, db *Tensor, pad, stride, dilation []int32) (conv *Layer, err error)

ConvolutionLayer creates a convolution layer

func (*Builder) CreateConvolutionWeights 

func (l *Builder) CreateConvolutionWeights(dims []int32) (w, dw, b, db *Tensor, err error)

CreateConvolutionWeights creates the weights and delta weights of a convolution layer

func (*Builder) CreateDeconvolutionWeights 

func (l *Builder) CreateDeconvolutionWeights(dims []int32) (w, dw, b, db *Tensor, err error)

CreateDeconvolutionWeights creates the weights and delta weights of a deconvolution layer

func (*Builder) CreateRandomTensor 

func (l *Builder) CreateRandomTensor(dims []int32, mean, std float32, seed uint64) (t *Tensor, err error)

CreateRandomTensor creates a random tensor

func (*Builder) CreateTensor 

func (l *Builder) CreateTensor(dims []int32) (t *Tensor, err error)

CreateTensor creates a tensor

func (*Builder) Dropout 

func (l *Builder) Dropout(id int64, dropoutpercent float32, seed uint64) (d *Layer, err error)

Dropout creates an Dropout layer

func (*Builder) FindBiasTensor 

func (l *Builder) FindBiasTensor(dims []int32) (b *Tensor, err error)

FindBiasTensor finds the bias tensor according to the dims

func (*Builder) GetHandle 

func (l *Builder) GetHandle() *Handle

GetHandle returns the handle

func (*Builder) PoolingLayer 

func (l *Builder) PoolingLayer(id int64, window, padding, stride []int32) (p *Layer, err error)

PoolingLayer creates a pooling layer with flags set in Builder

func (*Builder) ReverseConvolutionLayer 

func (l *Builder) ReverseConvolutionLayer(id int64, groupcount int32, w, dw, b, db *Tensor, pad, stride, dilation []int32) (rconv *Layer, err error)

ReverseConvolutionLayer creates a reverse convolution layer

type Classifier 

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

Classifier will take the outputs of a neural network and find the error of it. To be passed back to the rest of the network.

type ClassifierModule 

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

ClassifierModule is used to classify outputs

func CreateCustomLossLayer 

func CreateCustomLossLayer(id int64, b *Builder, l LossLayer) (m *ClassifierModule)

CreateCustomLossLayer creates a module that uses l

func CreateMSEClassifier 

func CreateMSEClassifier(id int64, bldr *Builder, x, dx, target *Tensor) (m *ClassifierModule, err error)

CreateMSEClassifier sets the mean squared error classifier

func CreateSoftMaxClassifier 

func CreateSoftMaxClassifier(id int64, bldr *Builder, x, dx, y, target *Tensor) (m *ClassifierModule, err error)

CreateSoftMaxClassifier will create a simple module with each of the convolution layers being in parallel.

func (*ClassifierModule) GetAverageBatchLoss 

func (m *ClassifierModule) GetAverageBatchLoss() float32

GetAverageBatchLoss gets the average batch loss

func (*ClassifierModule) GetTensorDX 

func (m *ClassifierModule) GetTensorDX() (dx *Tensor)

GetTensorDX returns set dx tensor

func (*ClassifierModule) GetTensorDY 

func (m *ClassifierModule) GetTensorDY() (dy *Tensor)

GetTensorDY returns set dy tensor

func (*ClassifierModule) GetTensorX 

func (m *ClassifierModule) GetTensorX() (x *Tensor)

GetTensorX returns set x tensor

func (*ClassifierModule) GetTensorY 

func (m *ClassifierModule) GetTensorY() (y *Tensor)

GetTensorY returns set y tensor

func (*ClassifierModule) ID 

func (m *ClassifierModule) ID() int64

ID returns the id set for the module

func (*ClassifierModule) Inference 

func (m *ClassifierModule) Inference() error

Inference does a forward propagation without calculating errors

func (*ClassifierModule) PerformError 

func (m *ClassifierModule) PerformError() error

PerformError does the output and error calculation of the previous layer of the network

func (*ClassifierModule) SetTensorDX 

func (m *ClassifierModule) SetTensorDX(dx *Tensor)

SetTensorDX sets dx tensor

func (*ClassifierModule) SetTensorDY 

func (m *ClassifierModule) SetTensorDY(dy *Tensor)

SetTensorDY sets dy tensor

func (*ClassifierModule) SetTensorX 

func (m *ClassifierModule) SetTensorX(x *Tensor)

SetTensorX sets x tensor

func (*ClassifierModule) SetTensorY 

func (m *ClassifierModule) SetTensorY(y *Tensor)

SetTensorY sets y tensor

func (*ClassifierModule) TestForward 

func (m *ClassifierModule) TestForward() error

TestForward does the testforward so that loss can be seen

type Comm 

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

Comm is a communicator

func CreateComms 

func CreateComms(hs []*Handle) (comm []*Comm, err error)

CreateComms creates Communicators for parallel processes.

type CompressionModule 

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

CompressionModule is a module that concats several layers together when doing the forward and backward passes

func CreateCompressionModule 

func CreateCompressionModule(id int64, bldr *Builder,
	batch, inputchannels int32,
	paralleloutputchans, spacialdims []int32,
	paddingoffset int32,
	falpha, fbeta float64) (m *CompressionModule, err error)

CreateCompressionModule will create a simple module with each of the convolution layers being in parallel. The parallel convolution will have the same hw, but the channels for each can be changed. The number of convolutions depends on the length of the channel array. Each convolution will have pad = ((dim-1)*d +1 + offset)/2. Offset is usually 0, but offset can be used to change if the output will be even or odd. This will make the output for each convolution equal in the spacial dims. The strides are stuck at 2.

This considers a stride of 2 spacial dims (hw) need to be odd. This will preform a deconvolution with the formula for the output tensor:

N= batch;

C = [neurons[0]+ ... +neurons[i]];

H,W and more (spacial dims) = ((input-1)/2) + 1

the dx tensor is zeroed before back propagation If multiple modules share the same input. Each module will need its own tensor as its own dx output. Those tensors need to be summed into the output dy tensor of module it got its x input tensor from

func (CompressionModule) Backward 

func (m CompressionModule) Backward() (err error)

Backward does the backward propagation

func (CompressionModule) FindOutputDims 

func (m CompressionModule) FindOutputDims() (dims []int32, err error)

FindOutputDims returns the output dims of the module

func (CompressionModule) Forward 

func (m CompressionModule) Forward() (err error)

Forward does the forward operation

func (CompressionModule) GetTensorDX 

func (m CompressionModule) GetTensorDX() (dx *Tensor)

GetTensorDX returns set dx tensor

func (CompressionModule) GetTensorDY 

func (m CompressionModule) GetTensorDY() (dy *Tensor)

GetTensorDY returns set dy tensor

func (CompressionModule) GetTensorX 

func (m CompressionModule) GetTensorX() (x *Tensor)

GetTensorX returns set x tensor

func (CompressionModule) GetTensorY 

func (m CompressionModule) GetTensorY() (y *Tensor)

GetTensorY returns set y tensor

func (CompressionModule) ID 

func (m CompressionModule) ID() int64

ID is the id

func (CompressionModule) Inference 

func (m CompressionModule) Inference() (err error)

Inference does the inference forward operation

func (CompressionModule) InitHiddenLayers 

func (m CompressionModule) InitHiddenLayers(rate, decay1, decay2 float32) (err error)

InitHiddenLayers will init the hidden layers. If

func (CompressionModule) InitWorkspace 

func (m CompressionModule) InitWorkspace() (err error)

InitWorkspace inits the hidden workspace

func (CompressionModule) SetTensorDX 

func (m CompressionModule) SetTensorDX(dx *Tensor)

SetTensorDX sets dx tensor

func (CompressionModule) SetTensorDY 

func (m CompressionModule) SetTensorDY(dy *Tensor)

SetTensorDY sets dy tensor

func (CompressionModule) SetTensorX 

func (m CompressionModule) SetTensorX(x *Tensor)

SetTensorX sets x tensor

func (CompressionModule) SetTensorY 

func (m CompressionModule) SetTensorY(y *Tensor)

SetTensorY sets y tensor

func (CompressionModule) Update 

func (m CompressionModule) Update(epoch int) (err error)

Update updates the weights of the hidden convolution layer

type Concat 

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

Concat does the concat operation

func CreateConcat 

func CreateConcat(h *Handle) (c *Concat, err error)

CreateConcat creates a concat operation handler

func (*Concat) Backward 

func (c *Concat) Backward() error

Backward Does backward with data flowing dest to srcs

func (*Concat) FindOutputDims 

func (c *Concat) FindOutputDims(srcs []*Tensor) (outputdims []int32, err error)

FindOutputDims finds the output dims

func (*Concat) FindOutputDimsfromInputDims 

func (c *Concat) FindOutputDimsfromInputDims(srcs [][]int32, frmt TensorFormat) (outputdims []int32, err error)

FindOutputDimsfromInputDims finds the input dims from output dims

func (*Concat) Forward 

func (c *Concat) Forward() error

Forward Does forward with data flowing srcs to dest

func (*Concat) SetDeltaDest 

func (c *Concat) SetDeltaDest(deltadest *Tensor)

SetDeltaDest sets the delta dest for back propagation

func (*Concat) SetDest 

func (c *Concat) SetDest(dest *Tensor)

SetDest sets the output dest

func (*Concat) SetInputDeltaSrcs 

func (c *Concat) SetInputDeltaSrcs(deltasrcs []*Tensor)

SetInputDeltaSrcs sets the delta srcs for back propagation

func (*Concat) SetInputSrcs 

func (c *Concat) SetInputSrcs(srcs []*Tensor)

SetInputSrcs sets the input srcs

type ConvolutionMode 

type ConvolutionMode struct {
	gocudnn.ConvolutionMode
}

ConvolutionMode struct wrapper for gocudnn.ConvolutionMode. Look up methods in gocudnn.

func (*ConvolutionMode) Convolution 

func (c *ConvolutionMode) Convolution() ConvolutionMode

Convolution sets and returns the Convolution flag

func (*ConvolutionMode) CrossCorrelation 

func (c *ConvolutionMode) CrossCorrelation() ConvolutionMode

CrossCorrelation sets and returns the CrossCorrelation flag

type DataType 

type DataType struct {
	gocudnn.DataType
}

DataType struct wrapper for gocudnn.Datatype. Look up methods in gocudnn.

func (*DataType) Double 

func (d *DataType) Double() DataType

Double sets and returns the Double flag

func (*DataType) Float 

func (d *DataType) Float() DataType

Float sets and returns the Float flag

func (*DataType) Half 

func (d *DataType) Half() DataType

Half sets and returns the Half flag

func (*DataType) Int32 

func (d *DataType) Int32() DataType

Int32 sets and returns the Int32 flag

func (*DataType) Int8 

func (d *DataType) Int8() DataType

Int8 sets and returns the Int8 flag

func (*DataType) Int8x32 

func (d *DataType) Int8x32() DataType

Int8x32 sets and returns the Int32 flag

func (*DataType) Int8x4 

func (d *DataType) Int8x4() DataType

Int8x4 sets and returns the Int32 flag

func (*DataType) UInt8 

func (d *DataType) UInt8() DataType

UInt8 sets and returns the Uint8 flag

func (*DataType) UInt8x4 

func (d *DataType) UInt8x4() DataType

UInt8x4 sets and returns the Int32 flag

type DecompressionModule 

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

DecompressionModule is a module that concats several layers together when doing the forward and backward passes

func CreateDecompressionModule 

func CreateDecompressionModule(id int64, bldr *Builder, batch, inputchannel int32, outputperparallellayer, spacialdims []int32, paddingoffset int32, falpha, fbeta float64) (m *DecompressionModule, err error)

CreateDecompressionModule will create a simple module with each of the deconvolution layers being in parallel. Deconvolution output channels is determined by the size of the neuron channels. The number of neurons needs to equal the input channels.

The parallel convolution will have the same hw, but the neuron channels for each can be changed. The number of convolutions depends on the length of the channel array. Each convolution will have pad = ((dim-1)/2) *dilation. This will make the output for each convolution equal in the spacial dims. The strides are stuck at 2.

This considers a stride of 2 spacial dims (hw) need to be odd. This will preform a deconvolution with the formula for the output tensor:

N= batch;

C = [neuronchannels[0]+ ... +neuronchannels[i]];

H,W and more (spacial dims) = 2*input - 1

the dx tensor is zeroed before back propagation If multiple modules share the same input. Each module will need its own tensor as its own dx output. Those tensors need to be summed into the output dy tensor of module it got its x input tensor from

func (DecompressionModule) Backward 

func (m DecompressionModule) Backward() (err error)

Backward does the backward propagation

func (DecompressionModule) FindOutputDims 

func (m DecompressionModule) FindOutputDims() (dims []int32, err error)

FindOutputDims returns the output dims of the module

func (DecompressionModule) Forward 

func (m DecompressionModule) Forward() (err error)

Forward does the forward operation

func (DecompressionModule) GetTensorDX 

func (m DecompressionModule) GetTensorDX() (dx *Tensor)

GetTensorDX returns set dx tensor

func (DecompressionModule) GetTensorDY 

func (m DecompressionModule) GetTensorDY() (dy *Tensor)

GetTensorDY returns set dy tensor

func (DecompressionModule) GetTensorX 

func (m DecompressionModule) GetTensorX() (x *Tensor)

GetTensorX returns set x tensor

func (DecompressionModule) GetTensorY 

func (m DecompressionModule) GetTensorY() (y *Tensor)

GetTensorY returns set y tensor

func (DecompressionModule) ID 

func (m DecompressionModule) ID() int64

ID is the id

func (DecompressionModule) Inference 

func (m DecompressionModule) Inference() (err error)

Inference does the inference forward operation

func (DecompressionModule) InitHiddenLayers 

func (m DecompressionModule) InitHiddenLayers(rate, decay1, decay2 float32) (err error)

InitHiddenLayers will init the hidden layers. If

func (DecompressionModule) InitWorkspace 

func (m DecompressionModule) InitWorkspace() (err error)

InitWorkspace inits the hidden workspace

func (DecompressionModule) SetTensorDX 

func (m DecompressionModule) SetTensorDX(dx *Tensor)

SetTensorDX sets dx tensor

func (DecompressionModule) SetTensorDY 

func (m DecompressionModule) SetTensorDY(dy *Tensor)

SetTensorDY sets dy tensor

func (DecompressionModule) SetTensorX 

func (m DecompressionModule) SetTensorX(x *Tensor)

SetTensorX sets x tensor

func (DecompressionModule) SetTensorY 

func (m DecompressionModule) SetTensorY(y *Tensor)

SetTensorY sets y tensor

func (DecompressionModule) Update 

func (m DecompressionModule) Update(epoch int) (err error)

Update updates the weights of the hidden convolution layer

type Device 

type Device struct {
	cudart.Device
	// contains filtered or unexported fields
}

Device is a gpu device

func GetDeviceList 

func GetDeviceList() (devices []Device, err error)

GetDeviceList gets a device from a list

func (Device) Num 

func (d Device) Num() int32

Num is the numerical id of the device

type Forwarder 

type Forwarder interface {
	Forward() error
}

Forwarder does the forward operation

type Handle 

type Handle struct {
	*cudnn.Handler
}

Handle handles the functions of the libraries used in gocunet

func CreateHandle 

func CreateHandle(w *Worker, d Device, seed uint64) (h *Handle)

CreateHandle creates a handle for gocunets

func CreateHandles 

func CreateHandles(ws []*Worker, ds []Device, seeds []uint64) []*Handle

CreateHandles creates parrallel handles. With there own workers. It also creates non blocking streams

func (*Handle) GetWorker 

func (h *Handle) GetWorker() *Worker

GetWorker returns the gocu.Worker.

type Layer 

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

Layer is a layer inside a network it holds inputs and outputs

func CreateOperationLayer 

func CreateOperationLayer(id int64, handle *Handle, op Operation) (l *Layer, err error)

CreateOperationLayer creates an operation layer

func (*Layer) Backward 

func (l *Layer) Backward() error

Backward performs the backward propagation

func (*Layer) ChangeBatchSize 

func (l *Layer) ChangeBatchSize(batchsize int)

ChangeBatchSize will change the batch size

func (*Layer) Forward 

func (l *Layer) Forward() error

Forward performs the forward propagation

func (*Layer) GetOutputDims 

func (l *Layer) GetOutputDims(input *Tensor) (output []int32, err error)

GetOutputDims gets the dims of the output tensor

func (*Layer) ID 

func (l *Layer) ID() int64

ID is the ID of the layer

func (*Layer) LoadTrainer 

func (l *Layer) LoadTrainer(handle *cudnn.Handler, batchsize int, trainers ...trainer.Trainer) error

LoadTrainer Loas the trainer to the layer

func (*Layer) SetBackwardScalars 

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

SetBackwardScalars sets backward scalars

func (*Layer) SetForwardScalars 

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

SetForwardScalars sets the forward scalars.

func (*Layer) SetIOs 

func (l *Layer) SetIOs(x, dx, y, dy *Tensor)

SetIOs sets the x,dx,y,dy used by the layer

func (*Layer) SetInputs 

func (l *Layer) SetInputs(x, dx *Tensor)

SetInputs sets the inputs

func (*Layer) SetOtherScalars 

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

SetOtherScalars sets other scalars that the layer might have scalars

func (*Layer) SetOutputs 

func (l *Layer) SetOutputs(y, dy *Tensor)

SetOutputs sets the outputs

func (*Layer) String 

func (l *Layer) String() string

func (*Layer) ToggleAllHiddenLayerValues 

func (l *Layer) ToggleAllHiddenLayerValues()

ToggleAllHiddenLayerValues toggles all hidden values on

func (*Layer) ToggleBiasPrint 

func (l *Layer) ToggleBiasPrint()

ToggleBiasPrint If layer contains bias.

func (*Layer) ToggleDBiasPrint 

func (l *Layer) ToggleDBiasPrint()

ToggleDBiasPrint If layer contains dbias

func (*Layer) ToggleDWPrint 

func (l *Layer) ToggleDWPrint()

ToggleDWPrint If layer contains delta weights

func (*Layer) ToggleWPrint 

func (l *Layer) ToggleWPrint()

ToggleWPrint If layer contains weights or hidden values it will toggle the printing

func (*Layer) Update 

func (l *Layer) Update(epoch int) error

Update updates weights if layer has them

type LossLayer 

type LossLayer interface {
	PerformError(x, dx, y, dy *layers.Tensor) (err error)
	Inference(x, y *layers.Tensor) (err error)
	TestForward(x, y, target *layers.Tensor) (err error)
	GetAverageBatchLoss() float32
}

LossLayer performs two functions. Be able to calculate loss, and to be able to calculate the inference forward.

type LossMode 

type LossMode int

LossMode is the flags for loss mode

type LossModeFlag 

type LossModeFlag struct {
}

LossModeFlag will return flags for LossMode These will be added over time.

func (LossModeFlag) Binary 

func (l LossModeFlag) Binary() LossMode

Binary is a loss mode

func (LossModeFlag) Huber 

func (l LossModeFlag) Huber() LossMode

Huber is a loss mode

type MathType 

type MathType struct {
	gocudnn.MathType
}

MathType is math type for tensor cores

func (*MathType) AllowConversion 

func (m *MathType) AllowConversion() MathType

AllowConversion sets and returns the AllowConversion flag

func (*MathType) Default 

func (m *MathType) Default() MathType

Default sets and returns the Default flag

func (*MathType) TensorOpMath 

func (m *MathType) TensorOpMath() MathType

TensorOpMath sets and returns the TensorOpMath flag

type Module 

type Module interface {
	ID() int64
	Forward() error
	Backward() error
	Update(counter int) error //counter can count updates or it can count epochs.  I found updates to work best.
	FindOutputDims() ([]int32, error)
	Inference() error
	InitHiddenLayers(rate, decay1, decay2 float32) (err error)
	InitWorkspace() (err error)
	GetTensorX() (x *Tensor)
	GetTensorDX() (dx *Tensor)
	GetTensorY() (y *Tensor)
	GetTensorDY() (dy *Tensor)
	SetTensorX(x *Tensor)
	SetTensorDX(dx *Tensor)
	SetTensorY(y *Tensor)
	SetTensorDY(dy *Tensor)
}

Module is a wrapper around a neural network or set of operations

type NanProp 

type NanProp struct {
	gocudnn.NANProp
}

NanProp struct wrapper for gocudnn.NanProp. Look up methods in gocudnn.

func (*NanProp) NotPropigate 

func (n *NanProp) NotPropigate() NanProp

NotPropigate sets and returns the NotPropigate flag

func (*NanProp) Propigate 

func (n *NanProp) Propigate() NanProp

Propigate sets and returns the Propigate flag

type NeutralModule 

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

NeutralModule is for nonsliding modules

func CreateSingleStridedModule 

func CreateSingleStridedModule(id int64, bldr *Builder,
	batch, inputchannels int32,
	paralleloutputchans, spacialdims []int32,
	paddingoffset int32,
	falpha, fbeta float64, strides, deconv bool) (m *NeutralModule, err error)

CreateSingleStridedModule creates a Module with stride set to 1. Since there are parallel convolution in this layer the input is shared between all of them. the dx tensor is zeroed before back propagation If multiple modules share the same input. Each module will need its own tensor as its own dx output. Those tensors need to be summed into the output dy tensor of module it got its x input tensor from

func (NeutralModule) Backward 

func (m NeutralModule) Backward() (err error)

Backward does the backward propagation

func (NeutralModule) FindOutputDims 

func (m NeutralModule) FindOutputDims() (dims []int32, err error)

FindOutputDims returns the output dims of the module

func (NeutralModule) Forward 

func (m NeutralModule) Forward() (err error)

Forward does the forward operation

func (NeutralModule) GetTensorDX 

func (m NeutralModule) GetTensorDX() (dx *Tensor)

GetTensorDX returns set dx tensor

func (NeutralModule) GetTensorDY 

func (m NeutralModule) GetTensorDY() (dy *Tensor)

GetTensorDY returns set dy tensor

func (NeutralModule) GetTensorX 

func (m NeutralModule) GetTensorX() (x *Tensor)

GetTensorX returns set x tensor

func (NeutralModule) GetTensorY 

func (m NeutralModule) GetTensorY() (y *Tensor)

GetTensorY returns set y tensor

func (NeutralModule) ID 

func (m NeutralModule) ID() int64

ID is the id

func (NeutralModule) Inference 

func (m NeutralModule) Inference() (err error)

Inference does the inference forward operation

func (NeutralModule) InitHiddenLayers 

func (m NeutralModule) InitHiddenLayers(rate, decay1, decay2 float32) (err error)

InitHiddenLayers will init the hidden layers. If

func (NeutralModule) InitWorkspace 

func (m NeutralModule) InitWorkspace() (err error)

InitWorkspace inits the hidden workspace

func (NeutralModule) SetTensorDX 

func (m NeutralModule) SetTensorDX(dx *Tensor)

SetTensorDX sets dx tensor

func (NeutralModule) SetTensorDY 

func (m NeutralModule) SetTensorDY(dy *Tensor)

SetTensorDY sets dy tensor

func (NeutralModule) SetTensorX 

func (m NeutralModule) SetTensorX(x *Tensor)

SetTensorX sets x tensor

func (NeutralModule) SetTensorY 

func (m NeutralModule) SetTensorY(y *Tensor)

SetTensorY sets y tensor

func (NeutralModule) Update 

func (m NeutralModule) Update(epoch int) (err error)

Update updates the weights of the hidden convolution layer

type Operation 

type Operation interface {
	Forward(handle *cudnn.Handler, x, dx, y, dy *layers.Tensor) error
	Inference(handle *cudnn.Handler, x, y *layers.Tensor) error
	Backward(handle *cudnn.Handler, x, dx, y, dy *Tensor) error
	UpdateWeights(handle *cudnn.Handler) error
	LoadTrainers(handle *cudnn.Handler, trainers ...trainer.Trainer) error
	TrainersNeeded() int
	SetOtherScalars(alpha, beta float64)
	SetForwardScalars(alpha, beta float64)
	SetBackwardScalars(alpha, beta float64)
	GetOutputDims(input *layers.Tensor) ([]int32, error)
}

Operation is a generic operation that a layer uses.

The forward and backward don't need to use all the x,dx,y,and dy, but they do need to be passed.

type OutputModule 

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

OutputModule is just a single single convolution before it goes into the loss function

func CreateOutputModule 

func CreateOutputModule(id int64, bldr *Builder, batch int32, fdims, pad, stride, dilation []int32, balpha, bbeta, falpha, fbeta float64) (m *OutputModule, err error)

CreateOutputModule creates an output module

func (*OutputModule) Backward 

func (m *OutputModule) Backward() error

Backward satisfies module interface

func (*OutputModule) FindOutputDims 

func (m *OutputModule) FindOutputDims() ([]int32, error)

FindOutputDims satisifies module interface

func (*OutputModule) Forward 

func (m *OutputModule) Forward() error

Forward satisfies module interface

func (*OutputModule) GetTensorDX 

func (m *OutputModule) GetTensorDX() (dx *Tensor)

GetTensorDX returns set dx tensor

func (*OutputModule) GetTensorDY 

func (m *OutputModule) GetTensorDY() (dy *Tensor)

GetTensorDY returns set dy tensor

func (*OutputModule) GetTensorX 

func (m *OutputModule) GetTensorX() (x *Tensor)

GetTensorX returns set x tensor

func (*OutputModule) GetTensorY 

func (m *OutputModule) GetTensorY() (y *Tensor)

GetTensorY returns set y tensor

func (*OutputModule) ID 

func (m *OutputModule) ID() int64

ID satisfies module interface

func (*OutputModule) Inference 

func (m *OutputModule) Inference() error

Inference satisfies module interface

func (*OutputModule) InitHiddenLayers 

func (m *OutputModule) InitHiddenLayers(rate, decay1, decay2 float32) (err error)

InitHiddenLayers will init the hidden operation

func (*OutputModule) InitWorkspace 

func (m *OutputModule) InitWorkspace() (err error)

InitWorkspace inits the workspace

func (*OutputModule) SetTensorDX 

func (m *OutputModule) SetTensorDX(dx *Tensor)

SetTensorDX sets dx tensor

func (*OutputModule) SetTensorDY 

func (m *OutputModule) SetTensorDY(dy *Tensor)

SetTensorDY sets dy tensor

func (*OutputModule) SetTensorX 

func (m *OutputModule) SetTensorX(x *Tensor)

SetTensorX sets x tensor

func (*OutputModule) SetTensorY 

func (m *OutputModule) SetTensorY(y *Tensor)

SetTensorY sets y tensor

func (*OutputModule) Update 

func (m *OutputModule) Update(epoch int) error

Update satisifies module interface

type PoolingMode 

type PoolingMode struct {
	gocudnn.PoolingMode
}

PoolingMode struct wrapper for gocudnn.PoolingMode. Look up methods in gocudnn.

func (*PoolingMode) AverageCountExcludePadding 

func (p *PoolingMode) AverageCountExcludePadding() PoolingMode

AverageCountExcludePadding sets and returns the AverageCountExcludePadding flag

func (*PoolingMode) AverageCountIncludePadding 

func (p *PoolingMode) AverageCountIncludePadding() PoolingMode

AverageCountIncludePadding sets and returns the AverageCountIncludePadding flag

func (*PoolingMode) Max 

func (p *PoolingMode) Max() PoolingMode

Max sets and returns the Max flag

func (*PoolingMode) MaxDeterministic 

func (p *PoolingMode) MaxDeterministic() PoolingMode

MaxDeterministic sets and returns the MaxDeterministic flag

type ReverseConcat 

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

ReverseConcat is just a simple solution to split a source into multiple dests. If the source channel is not divisible by the number of dests then the remainder of the sources channels will be set into the last dest.

func CreateReverseConcat 

func CreateReverseConcat(h *Handle) (c *ReverseConcat, err error)

CreateReverseConcat creates a reverse concat

func (*ReverseConcat) Backward 

func (c *ReverseConcat) Backward() error

Backward Does backward with data flowing dest to srcs

func (*ReverseConcat) FindOutputDims 

func (c *ReverseConcat) FindOutputDims(Source *Tensor, ndests int32) (outputdims [][]int32, err error)

FindOutputDims finds the output dims for the dests

func (*ReverseConcat) FindOutputDimsfromInputDims 

func (c *ReverseConcat) FindOutputDimsfromInputDims(src []int32, ndests int32, frmt TensorFormat) (destdims [][]int32, err error)

FindOutputDimsfromInputDims finds the input dims from output dims. Last dest will get + the remainder for overflow or just the remainder of underflow

func (*ReverseConcat) Forward 

func (c *ReverseConcat) Forward() error

Forward Does forward with data flowing srcs to dest

func (*ReverseConcat) SetInputDeltaSource 

func (c *ReverseConcat) SetInputDeltaSource(deltasrc *Tensor)

SetInputDeltaSource sets the delta src for back propagation

func (*ReverseConcat) SetInputSource 

func (c *ReverseConcat) SetInputSource(src *Tensor)

SetInputSource sets the input source

func (*ReverseConcat) SetOutputDeltaDests 

func (c *ReverseConcat) SetOutputDeltaDests(deltadests []*Tensor)

SetOutputDeltaDests sets the delta dests for back propagation

func (*ReverseConcat) SetOutputDests 

func (c *ReverseConcat) SetOutputDests(dests []*Tensor)

SetOutputDests sets the output dests

type SimpleModuleNetwork 

type SimpleModuleNetwork struct {
	Id         int64             `json:"id,omitempty"`
	C          *Concat           `json:"c,omitempty"`
	Modules    []Module          `json:"modules,omitempty"`
	Output     *OutputModule     `json:"output,omitempty"`
	Classifier *ClassifierModule `json:"classifier,omitempty"`

	Rate, Decay1, Decay2 float32
	// contains filtered or unexported fields
}

SimpleModuleNetwork is a simple module network

func CreateSimpleModuleNetwork 

func CreateSimpleModuleNetwork(id int64, b *Builder) (smn *SimpleModuleNetwork)

CreateSimpleModuleNetwork a simple module network

func (*SimpleModuleNetwork) Backward 

func (m *SimpleModuleNetwork) Backward() (err error)

Backward does a forward without a concat

func (*SimpleModuleNetwork) BackwardCustom 

func (m *SimpleModuleNetwork) BackwardCustom(backward func() error) (err error)

BackwardCustom does a custom backward function

func (*SimpleModuleNetwork) FindOutputDims 

func (m *SimpleModuleNetwork) FindOutputDims() (dims []int32, err error)

FindOutputDims satisifis the Module interface

Have to run (m *SimpleModuleNetwork)SetTensorX(). If module network requres backpropdata to go to another module network. Then also run (m *SimpleModuleNetwork)SetTensorDX()

func (*SimpleModuleNetwork) Forward 

func (m *SimpleModuleNetwork) Forward() (err error)

Forward does a forward without a concat

func (*SimpleModuleNetwork) ForwardCustom 

func (m *SimpleModuleNetwork) ForwardCustom(forward func() error) (err error)

ForwardCustom does a custom forward function

func (*SimpleModuleNetwork) GetLoss 

func (m *SimpleModuleNetwork) GetLoss() float32

GetLoss returns the loss found.

func (*SimpleModuleNetwork) GetTensorDX 

func (m *SimpleModuleNetwork) GetTensorDX() *Tensor

GetTensorDX Gets dx tensor

func (*SimpleModuleNetwork) GetTensorDY 

func (m *SimpleModuleNetwork) GetTensorDY() *Tensor

GetTensorDY Gets dy tensor

func (*SimpleModuleNetwork) GetTensorX 

func (m *SimpleModuleNetwork) GetTensorX() *Tensor

GetTensorX Gets x tensor

func (*SimpleModuleNetwork) GetTensorY 

func (m *SimpleModuleNetwork) GetTensorY() *Tensor

GetTensorY Gets y tensor

func (*SimpleModuleNetwork) ID 

func (m *SimpleModuleNetwork) ID() int64

ID satisfies Module interface

func (*SimpleModuleNetwork) Inference 

func (m *SimpleModuleNetwork) Inference() (err error)

Inference does a forward without a concat

func (*SimpleModuleNetwork) InitHiddenLayers 

func (m *SimpleModuleNetwork) InitHiddenLayers(rate, decay1, decay2 float32) (err error)

InitHiddenLayers satisfies the Module interface

func (*SimpleModuleNetwork) InitWorkspace 

func (m *SimpleModuleNetwork) InitWorkspace() (err error)

InitWorkspace inits workspace

func (*SimpleModuleNetwork) SetMSEClassifier 

func (m *SimpleModuleNetwork) SetMSEClassifier() (err error)

SetMSEClassifier needs to be made

func (*SimpleModuleNetwork) SetModules 

func (m *SimpleModuleNetwork) SetModules(modules []Module)

SetModules sets modules

func (*SimpleModuleNetwork) SetSoftMaxClassifier 

func (m *SimpleModuleNetwork) SetSoftMaxClassifier() (err error)

SetSoftMaxClassifier sets the classifier module it should be added last. Should be ran after OutputModule is set

func (*SimpleModuleNetwork) SetTensorDX 

func (m *SimpleModuleNetwork) SetTensorDX(dx *Tensor)

SetTensorDX sets dx tensor

func (*SimpleModuleNetwork) SetTensorDY 

func (m *SimpleModuleNetwork) SetTensorDY(dy *Tensor)

SetTensorDY sets dy tensor

func (*SimpleModuleNetwork) SetTensorX 

func (m *SimpleModuleNetwork) SetTensorX(x *Tensor)

SetTensorX sets x tensor

func (*SimpleModuleNetwork) SetTensorY 

func (m *SimpleModuleNetwork) SetTensorY(y *Tensor)

SetTensorY sets y tensor

func (*SimpleModuleNetwork) TestForward 

func (m *SimpleModuleNetwork) TestForward() (err error)

TestForward does the forward prop but it still calculates loss for testing

func (*SimpleModuleNetwork) Update 

func (m *SimpleModuleNetwork) Update(counter int) (err error)

Update updates the hidden weights Update can count epochs or updates. I found counting updates works the best.

type SoftmaxAlgo 

type SoftmaxAlgo struct {
	gocudnn.SoftMaxAlgorithm
}

SoftmaxAlgo determins what algo to use for softmax

func (*SoftmaxAlgo) Accurate 

func (s *SoftmaxAlgo) Accurate() SoftmaxAlgo

Accurate sets and returns the Accurate flag

func (*SoftmaxAlgo) Fast 

func (s *SoftmaxAlgo) Fast() SoftmaxAlgo

Fast sets and returns the Fast flag

func (*SoftmaxAlgo) Log 

func (s *SoftmaxAlgo) Log() SoftmaxAlgo

Log sets and returns the Log flag

type SoftmaxMode 

type SoftmaxMode struct {
	gocudnn.SoftMaxMode
}

SoftmaxMode determins what mode to use for softmax

func (*SoftmaxMode) Channel 

func (s *SoftmaxMode) Channel() SoftmaxMode

Channel sets and returns the Channel flag

func (*SoftmaxMode) Instance 

func (s *SoftmaxMode) Instance() SoftmaxMode

Instance sets and returns the Instance flag

type Stream 

type Stream struct {
	*cudart.Stream
}

Stream is a stream for gpu instructions

func CreateStream 

func CreateStream() (s *Stream, err error)

CreateStream creates a stream

type Tensor 

type Tensor struct {
	*layers.Tensor
}

Tensor is contains 2 tensors the x and dx. Input IOs will contain only the X tensor.

type TensorFormat 

type TensorFormat struct {
	gocudnn.TensorFormat
}

TensorFormat struct wrapper for gocudnn.TensorFormat. Look up methods in gocudnn.

func (*TensorFormat) NCHW 

func (t *TensorFormat) NCHW() TensorFormat

NCHW sets and returns the NCHW flag

func (*TensorFormat) NCHWvectC 

func (t *TensorFormat) NCHWvectC() TensorFormat

NCHWvectC sets and returns the NHWC flag

func (*TensorFormat) NHWC 

func (t *TensorFormat) NHWC() TensorFormat

NHWC sets and returns the NHWC flag

func (*TensorFormat) Unknown 

func (t *TensorFormat) Unknown() TensorFormat

Unknown sets and returns the NHWC flag

type Updater 

type Updater interface {
	Update() error
}

Updater does the update interface

type VanillaModule 

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

VanillaModule has a convolution and an activation

func CreateVanillaModule 

func CreateVanillaModule(id int64, bldr *Builder, batch int32, fdims, pad, stride, dilation []int32, balpha, bbeta, falpha, fbeta float64) (m *VanillaModule, err error)

CreateVanillaModule creates an output module

func (*VanillaModule) Backward 

func (m *VanillaModule) Backward() error

Backward satisfies module interface

func (*VanillaModule) FindOutputDims 

func (m *VanillaModule) FindOutputDims() ([]int32, error)

FindOutputDims satisifies module interface

func (*VanillaModule) Forward 

func (m *VanillaModule) Forward() error

Forward satisfies module interface

func (*VanillaModule) GetTensorDX 

func (m *VanillaModule) GetTensorDX() (dx *Tensor)

GetTensorDX returns set dx tensor

func (*VanillaModule) GetTensorDY 

func (m *VanillaModule) GetTensorDY() (dy *Tensor)

GetTensorDY returns set dy tensor

func (*VanillaModule) GetTensorX 

func (m *VanillaModule) GetTensorX() (x *Tensor)

GetTensorX returns set x tensor

func (*VanillaModule) GetTensorY 

func (m *VanillaModule) GetTensorY() (y *Tensor)

GetTensorY returns set y tensor

func (*VanillaModule) ID 

func (m *VanillaModule) ID() int64

ID satisfies module interface

func (*VanillaModule) Inference 

func (m *VanillaModule) Inference() error

Inference satisfies module interface

func (*VanillaModule) InitHiddenLayers 

func (m *VanillaModule) InitHiddenLayers(rate, decay1, decay2 float32) (err error)

InitHiddenLayers will init the hidden operation

func (*VanillaModule) InitWorkspace 

func (m *VanillaModule) InitWorkspace() (err error)

InitWorkspace inits the workspace

func (*VanillaModule) SetTensorDX 

func (m *VanillaModule) SetTensorDX(dx *Tensor)

SetTensorDX sets dx tensor

func (*VanillaModule) SetTensorDY 

func (m *VanillaModule) SetTensorDY(dy *Tensor)

SetTensorDY sets dy tensor

func (*VanillaModule) SetTensorX 

func (m *VanillaModule) SetTensorX(x *Tensor)

SetTensorX sets x tensor

func (*VanillaModule) SetTensorY 

func (m *VanillaModule) SetTensorY(y *Tensor)

SetTensorY sets y tensor

func (*VanillaModule) Update 

func (m *VanillaModule) Update(epoch int) error

Update satisifies module interface

type Worker 

type Worker struct {
	*gocu.Worker
}

Worker is a wrapper for *gocu.Worker it assigns a locked host thread to a device. A device can have more than one worker.

func CreateWorker 

func CreateWorker(d Device) (w *Worker)

CreateWorker assigns a locked host thread to a device.

Source Files

View all Source files

Directories

Path Synopsis
Package devices is kind of a pipe dream.
 Package devices is kind of a pipe dream.
gpu/nvidia
gpu/nvidia/cuda/cudart
gpu/nvidia/cudnn
Package cudnn takes the descriptors from gocudnn which is from cudnn and seperates them into seperate packages.
 Package cudnn takes the descriptors from gocudnn which is from cudnn and seperates them into seperate packages.
gpu/nvidia/cudnn/activation
gpu/nvidia/cudnn/batchnorm
gpu/nvidia/cudnn/convolution
gpu/nvidia/cudnn/convolution/testmain
gpu/nvidia/cudnn/deconvolution
gpu/nvidia/cudnn/dropout
gpu/nvidia/cudnn/pool
gpu/nvidia/cudnn/reduce
gpu/nvidia/cudnn/softmax
Package softmax uses the softmax functions from gocudnn which is from cudnn.
 Package softmax uses the softmax functions from gocudnn which is from cudnn.
gpu/nvidia/cudnn/tensor
Package tensor is used to make tensors by using gocudnn.
 Package tensor is used to make tensors by using gocudnn.
gpu/nvidia/cudnn/transform
gpu/nvidia/custom/reshapes
gpu/nvidia/custom/swapper
gpu/nvidia/custom/xloss
gpu/nvidia/jpeg
gpu/nvidia/nonpp
gpu/nvidia/nvutil
Package nvutil are functions that use the other nvidia packages and allows them to be used with each other
 Package nvutil are functions that use the other nvidia packages and allows them to be used with each other
gpu/nvidia/nvutil/testing
Package layers contains shared things between layers.
 Package layers contains shared things between layers.
activation
batchnorm
cnn
Package cnn contains structs and methods used to do forward, and backward operations for convolution layers
 Package cnn contains structs and methods used to do forward, and backward operations for convolution layers
cnntranspose
dropout
pooling
reshape
softmax
transform
gansloss
test
mnist/dfuncs
mnist/mnistgpu
mnistauto
mnistauto/networks
mnistdcauto
mnistdcauto/dcnetworks
mnistgputest2
mnisttoroman
mnisttoroman/roman
parallelgpus
parallelgpus/neuralhelpers
Package trainer is a package that is used for training networks.
 Package trainer is a package that is used for training networks.
ui
gpuperformance
plot
plot/test
uihelper
filing
hwperf
imaging
outputdim
outputdim/debugmain
outputdim/dimlayer

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.