package gorgonia

import "gorgonia.org/gorgonia"

Package gorgonia is a library that helps facilitate machine learning in Go. Write and evaluate mathematical equations involving multidimensional arrays easily. Do differentiation with them just as easily.

g := NewGraph()

var x, y, z *Node
var err error

// define the expression
x = NewScalar(g, Float64, WithName("x"))
y = NewScalar(g, Float64, WithName("y"))
if z, err = Add(x, y); err != nil {
    log.Fatal(err)
}

// set initial values then run
Let(x, 2.0)
Let(y, 2.5)

// by default, LispMachine performs forward mode and backwards mode execution
m := NewLispMachine(g)
defer m.Close()
if err = m.RunAll(); err != nil {
    log.Fatal(err)
}

fmt.Printf("z: %v\n", z.Value())

if xgrad, err := x.Grad(); err == nil {
    fmt.Printf("dz/dx: %v\n", xgrad)
}

if ygrad, err := y.Grad(); err == nil {
    fmt.Printf("dz/dy: %v\n", ygrad)
}

z: 4.5
dz/dx: 1
dz/dy: 1

z = x + y

g := NewGraph()

var x, y, z *Node
var err error

// define the expression
x = NewScalar(g, Float64, WithName("x"))
y = NewScalar(g, Float64, WithName("y"))
if z, err = Add(x, y); err != nil {
    log.Fatal(err)
}

// create a VM to run the program on
machine := NewTapeMachine(g)
defer machine.Close()

// set initial values then run
Let(x, 2.0)
Let(y, 2.5)
if err = machine.RunAll(); err != nil {
    log.Fatal(err)
}

fmt.Printf("%v", z.Value())

4.5

xV, yV, bs := prep()
concurrentTraining(xV, yV, bs, epochs)

fmt.Printf("x:\n%1.1v", xV)
fmt.Printf("y:\n%1.1v", yV)

x:
⎡-0.0003     0.01     0.04     0.09      0.2⎤
⎢-0.0003     0.01     0.04     0.09      0.2⎥
⎢-0.0003     0.01     0.04     0.09      0.2⎥
⎢-0.0003     0.01     0.04     0.09      0.2⎥
.
.
.
⎢-0.0003     0.01     0.04     0.09      0.2⎥
⎢-0.0003     0.01     0.04     0.09      0.2⎥
⎢-0.0003     0.01     0.04     0.09      0.2⎥
⎣-0.0003     0.01     0.04     0.09      0.2⎦
y:
[0.3  0.3  0.3  0.3  ... 0.3  0.3  0.3  0.3]

// Lift
g := NewGraph()
x := NewMatrix(g, Float32, WithShape(2, 3), WithInit(RangedFrom(0)), WithName("a"))
y := NewMatrix(g, Float32, WithShape(3, 2), WithInit(ValuesOf(float32(2))), WithName("b"))
z := NewMatrix(g, Float32, WithShape(2, 1), WithInit(Zeroes()), WithName("bias"))
wrong := NewMatrix(g, Float64, WithShape(2, 3), WithInit(RangedFrom(0)), WithName("wrong"))

// Different LiftXXX functions exist for different API signatures
// A good way to do this is to have some instantiated functions at the top level of the package
mul := Lift2(Mul)
add := Lift2(Add)
addB := Lift2Broadcast(BroadcastAdd)
sq := Lift1(Square)
sm := Lift1Axial(SoftMax)

nn := sm(sq(addB(mul(x, y), z, nil, []byte{1}))) // OK
nnPlusWrong := add(nn, wrong)                    // Wrong types. Will Error
fmt.Printf("nn: %v\nAn error occurs: %v\n", nn, nnPlusWrong.Err())

// Must()
h := NewGraph()
a := NewMatrix(h, Float32, WithShape(2, 3), WithInit(RangedFrom(0)), WithName("a"))
b := NewMatrix(h, Float32, WithShape(3, 2), WithInit(ValuesOf(float32(2))), WithName("b"))
c := NewMatrix(h, Float32, WithShape(2, 1), WithInit(RangedFrom(0)), WithName("c"))
wrong2 := NewMatrix(h, Float64, WithShape(2, 3), WithInit(RangedFrom(0)), WithName("wrong"))

// This is OK
nn2 := Must(SoftMax(
    Must(Square(
        Must(BroadcastAdd(
            Must(Mul(a, b)),
            c,
            nil, []byte{1},
        )),
    )),
))

fmt.Printf("nn2: %v\n", nn2)

defer func() {
    if r := recover(); r != nil {
        fmt.Printf("An error occurs (caught by recover()): %v\n", r)
    }
}()

nn2PlusWrong := Must(Add(nn2, wrong2))
_ = nn2PlusWrong

nn: Softmax{-1, false}()(%9) :: Matrix float32
An error occurs: Type inference error. Op: + false. Children: [Matrix float32, Matrix float64], OpType:Matrix a → Matrix a → Matrix a: Unable to unify while inferring type of + false: Unification Fail: float64 ~ float32 cannot be unified
nn2: Softmax{-1, false}()(%9) :: Matrix float32
An error occurs (caught by recover()): Type inference error. Op: + false. Children: [Matrix float32, Matrix float64], OpType:Matrix a → Matrix a → Matrix a: Unable to unify while inferring type of + false: Unification Fail: float64 ~ float32 cannot be unified

g := NewGraph()
a := NodeFromAny(g, tensor.New(tensor.WithShape(2, 3), tensor.WithBacking([]float64{1, 2, 3, 4, 5, 6})))
m1, _ := Mean(a, 1)
m2, _ := KeepDims(a, false, func(a *Node) (*Node, error) { return Mean(a, 1) })
m3, _ := Mean(a, 0)
m4, _ := KeepDims(a, true, func(a *Node) (*Node, error) { return Mean(a, 0) })
m5, _ := KeepDims(a, true, func(a *Node) (*Node, error) { return Mean(a) })

// these reads are necessary as the VM may feel free to clobber the underlying data.
// e.g. if m1.Value() is used in the print statement below, the answer will be wrong.
// This is because before the VM executes the operations, a check is done to see if unsafe
// operations may be done. Unsafe operations are useful in saving memory.
// In this example, Reshape can be unsafely done if no other node is "using" m1,
// so m1.Value() will have its shape clobbered. Thus if m1.Value() is read after the VM has run,
// there is no guarantee that the data is correct. The only way around this is to "use" m1, by the Read() function.
var m1v, m2v, m3v, m4v Value
Read(m1, &m1v)
Read(m2, &m2v)
Read(m3, &m3v)
Read(m4, &m4v)

vm := NewTapeMachine(g)
if err := vm.RunAll(); err != nil {
    panic(err)
}

fmt.Printf("a:\n%v\n", a.Value())
fmt.Printf("m1 (shape: %v):\n%v\n", m1.Value().Shape(), m1v)
fmt.Printf("m2 (shape: %v):\n%v\n", m2.Value().Shape(), m2v)
fmt.Printf("m3 (shape: %v):\n%v\n", m3.Value().Shape(), m3v)
fmt.Printf("m4 (shape: %v):\n%v\n", m4.Value().Shape(), m4v)
fmt.Printf("m5 (shape: %v):\n%v\n", m5.Value().Shape(), m5.Value())

a:
⎡1  2  3⎤
⎣4  5  6⎦

m1 (shape: (2)):
[2  5]
m2 (shape: (2, 1)):
C[2  5]
m3 (shape: (3)):
[2.5  3.5  4.5]
m4 (shape: (1, 3)):
R[2.5  3.5  4.5]
m5 (shape: (1, 1)):
[[3.5]]

y = mx + c

package main

import (
    "fmt"
    "log"
    "math/rand"
    "runtime"

    . "gorgonia.org/gorgonia"
    "gorgonia.org/tensor"
)

const (
    vecSize = 1000000
)

// manually generate a fake dataset which is y=2x+random
func xy(dt tensor.Dtype) (x tensor.Tensor, y tensor.Tensor) {
    var xBack, yBack interface{}
    switch dt {
    case Float32:
        xBack = tensor.Range(tensor.Float32, 1, vecSize+1).([]float32)
        yBackC := tensor.Range(tensor.Float32, 1, vecSize+1).([]float32)

        for i, v := range yBackC {
            yBackC[i] = v*2 + rand.Float32()
        }
        yBack = yBackC
    case Float64:
        xBack = tensor.Range(tensor.Float64, 1, vecSize+1).([]float64)
        yBackC := tensor.Range(tensor.Float64, 1, vecSize+1).([]float64)

        for i, v := range yBackC {
            yBackC[i] = v*2 + rand.Float64()
        }
        yBack = yBackC
    }

    x = tensor.New(tensor.WithBacking(xBack), tensor.WithShape(vecSize))
    y = tensor.New(tensor.WithBacking(yBack), tensor.WithShape(vecSize))
    return
}

func random(dt tensor.Dtype) interface{} {
    rand.Seed(13370)
    switch dt {
    case tensor.Float32:
        return rand.Float32()
    case tensor.Float64:
        return rand.Float64()
    default:
        panic("Unhandled dtype")
    }
}

func linregSetup(Float tensor.Dtype) (m, c *Node, machine VM) {
    var xT, yT Value
    xT, yT = xy(Float)

    g := NewGraph()
    x := NewVector(g, Float, WithShape(vecSize), WithName("x"), WithValue(xT))
    y := NewVector(g, Float, WithShape(vecSize), WithName("y"), WithValue(yT))
    m = NewScalar(g, Float, WithName("m"), WithValue(random(Float)))
    c = NewScalar(g, Float, WithName("c"), WithValue(random(Float)))

    pred := Must(Add(Must(Mul(x, m)), c))
    se := Must(Square(Must(Sub(pred, y))))
    cost := Must(Mean(se))

    if _, err := Grad(cost, m, c); err != nil {
        log.Fatalf("Failed to backpropagate: %v", err)
    }

    // machine := NewLispMachine(g)  // you can use a LispMachine, but it'll be VERY slow.
    machine = NewTapeMachine(g, BindDualValues(m, c))
    return m, c, machine
}

func linregRun(m, c *Node, machine VM, iter int, autoCleanup bool) (retM, retC Value) {
    if autoCleanup {
        defer machine.Close()
    }
    model := []ValueGrad{m, c}
    solver := NewVanillaSolver(WithLearnRate(0.001), WithClip(5)) // good idea to clip

    if CUDA {
        runtime.LockOSThread()
        defer runtime.UnlockOSThread()
    }
    var err error
    for i := 0; i < iter; i++ {
        if err = machine.RunAll(); err != nil {
            fmt.Printf("Error during iteration: %v: %v\n", i, err)
            break
        }

        if err = solver.Step(model); err != nil {
            log.Fatal(err)
        }

        machine.Reset() // Reset is necessary in a loop like this
    }
    return m.Value(), c.Value()

}

func linearRegression(Float tensor.Dtype, iter int) (retM, retC Value) {
    defer runtime.GC()
    m, c, machine := linregSetup(Float)
    return linregRun(m, c, machine, iter, true)
}

// Linear Regression Example
//
// The formula for a straight line is
//		y = mx + c
// We want to find an `m` and a `c` that fits the equation well. We'll do it in both float32 and float64 to showcase the extensibility of Gorgonia
func main() {
    var m, c Value
    // Float32
    m, c = linearRegression(Float32, 500)
    fmt.Printf("float32: y = %3.3fx + %3.3f\n", m, c)

    // Float64
    m, c = linearRegression(Float64, 500)
    fmt.Printf("float64: y = %3.3fx + %3.3f\n", m, c)

}

xV, yV, _ := prep()
nonConcurrentTraining(xV, yV, epochs)

fmt.Printf("x:\n%1.1v", xV)
fmt.Printf("y:\n%1.1v", yV)

x:
⎡-0.0003     0.01     0.04     0.09      0.2⎤
⎢-0.0003     0.01     0.04     0.09      0.2⎥
⎢-0.0003     0.01     0.04     0.09      0.2⎥
⎢-0.0003     0.01     0.04     0.09      0.2⎥
.
.
.
⎢-0.0003     0.01     0.04     0.09      0.2⎥
⎢-0.0003     0.01     0.04     0.09      0.2⎥
⎢-0.0003     0.01     0.04     0.09      0.2⎥
⎣-0.0003     0.01     0.04     0.09      0.2⎦
y:
[0.3  0.3  0.3  0.3  ... 0.3  0.3  0.3  0.3]

g := NewGraph()

var x, y, z *Node
var err error

// define the expression
x = NewScalar(g, Float64, WithName("x"))
y = NewScalar(g, Float64, WithName("y"))
if z, err = Add(x, y); err != nil {
    log.Fatal(err)
}

// symbolically differentiate z with regards to x and y
// this adds the gradient nodes to the graph g
var grads Nodes
if grads, err = Grad(z, x, y); err != nil {
    log.Fatal(err)
}

// create a VM to run the program on
machine := NewTapeMachine(g)
defer machine.Close()

// set initial values then run
Let(x, 2.0)
Let(y, 2.5)
if err = machine.RunAll(); err != nil {
    log.Fatal(err)
}

fmt.Printf("z: %v\n", z.Value())
if xgrad, err := x.Grad(); err == nil {
    fmt.Printf("dz/dx: %v | %v\n", xgrad, grads[0].Value())
}

if ygrad, err := y.Grad(); err == nil {
    fmt.Printf("dz/dy: %v | %v\n", ygrad, grads[1].Value())
}

z: 4.5
dz/dx: 1 | 1
dz/dy: 1 | 1

Index ¶

• Constants
• Variables
• func BatchNorm(x, scale, bias *Node, momentum, epsilon float64) (retVal, γ, β *Node, op *BatchNormOp, err error)
• func BatchNorm1d(x, scale, bias *Node, momentum, epsilon float64) (retVal, γ, β *Node, op *BatchNormOp, err error)
• func Binomial32(trials, prob float64, s ...int) []float32
• func Binomial64(trials, prob float64, s ...int) []float64
• func Broadcast(a, b *Node, pattern BroadcastPattern) (*Node, *Node, error)
• func CheckOne(in Input) error
• func Compile(g *ExprGraph) (prog *program, locMap map[*Node]register, err error)
• func CompileFunction(g *ExprGraph, inputs, outputs Nodes) (prog *program, locMap map[*Node]register, err error)
• func DebugDerives()
• func DimSizersToShapes(ds []DimSizer) ([]tensor.Shape, error)
• func DontDebugDerives()
• func Err(e error) gErr
• func FmtNodeMap(m interface{}) mapFmt
• func Gaussian32(mean, stdev float64, s ...int) []float32
• func Gaussian64(mean, stdev float64, s ...int) []float64
• func GlorotEtAlN32(gain float64, s ...int) []float32
• func GlorotEtAlN64(gain float64, s ...int) []float64
• func GlorotEtAlU32(gain float64, s ...int) []float32
• func GlorotEtAlU64(gain float64, s ...int) []float64
• func GraphCollisionStats() (int, int, int)
• func HeEtAlN64(gain float64, s ...int) []float64
• func HeEtAlU64(gain float64, s ...int) []float64
• func Let(n *Node, be interface{}) error
• func Lift1(fn func(a *Node) (*Node, error)) func(a Input) Result
• func Lift1Axial(fn func(a *Node, axes ...int) (*Node, error)) func(a Input, axes ...int) Result
• func Lift2(fn func(a, b *Node) (*Node, error)) func(a, b Input) Result
• func Lift2Broadcast(fn func(a, b *Node, pat1, pat2 []byte) (*Node, error)) func(a, b Input, pat1, pat2 []byte) Result
• func NewLispMachine(g *ExprGraph, opts ...VMOpt) *lispMachine
• func NewTapeMachine(g *ExprGraph, opts ...VMOpt) *tapeMachine
• func ReturnNode(n *Node)
• func ReturnType(t hm.Type)
• func S(start int, opt ...int) tensor.Slice
• func SetDerivOf(deriv, of *Node)
• func SetOptimizationLevel(i int)
• func TransformResult(ins ...Input) func(a Input, err error) Result
• func TypeOf(v Value) hm.Type
• func Uniform32(low, high float64, s ...int) []float32
• func Uniform64(low, high float64, s ...int) []float64
• func UnsafeLet(n *Node, be interface{}) error
• func Use(b BLAS)
• func UseNonStable()
• func UseStabilization()
• func ValueClose(a, b Value) bool
• func ValueEq(a, b Value) bool
• func WalkGraph(start *Node) <-chan *Node
• func WithGraphName(name string) graphconopt
• type ADOp
• type AdaGradSolver
• func NewAdaGradSolver(opts ...SolverOpt) *AdaGradSolver
• func (s *AdaGradSolver) Step(model []ValueGrad) (err error)
• type AdamSolver
• func NewAdamSolver(opts ...SolverOpt) *AdamSolver
• func (s *AdamSolver) Step(model []ValueGrad) (err error)
• type Arena
• type AutoDiffError
• func (err AutoDiffError) Error() string
• type B
• func (v *B) Data() interface{}
• func (v *B) Dtype() tensor.Dtype
• func (v *B) Format(s fmt.State, c rune)
• func (v *B) MemSize() uintptr
• func (v *B) Pointer() unsafe.Pointer
• func (v *B) Shape() tensor.Shape
• func (v *B) Size() int
• func (v *B) Uintptr() uintptr
• type BLAS
• func WhichBLAS() BLAS
• type BarzilaiBorweinSolver
• func NewBarzilaiBorweinSolver(opts ...SolverOpt) *BarzilaiBorweinSolver
• func (s *BarzilaiBorweinSolver) Step(model []ValueGrad) (err error)
• type BatchNormOp
• func (op *BatchNormOp) Arity() int
• func (op *BatchNormOp) CallsExtern() bool
• func (op *BatchNormOp) DiffWRT(inputs int) []bool
• func (op *BatchNormOp) Do(values ...Value) (retVal Value, err error)
• func (op *BatchNormOp) DoDiff(ctx ExecutionContext, inputs Nodes, output *Node) error
• func (op *BatchNormOp) Hashcode() uint32
• func (op *BatchNormOp) InferShape(ns ...DimSizer) (tensor.Shape, error)
• func (op *BatchNormOp) OverwritesInput() int
• func (op *BatchNormOp) Reset() error
• func (op *BatchNormOp) ReturnsPtr() bool
• func (op *BatchNormOp) SetTesting()
• func (op *BatchNormOp) SetTraining()
• func (op *BatchNormOp) String() string
• func (op *BatchNormOp) SymDiff(inputs Nodes, output *Node, grad *Node) (retVal Nodes, err error)
• func (op *BatchNormOp) Type() hm.Type
• func (op *BatchNormOp) UsePreallocDo(prealloc Value, inputs ...Value) (retVal Value, err error)
• func (op *BatchNormOp) WriteHash(h hash.Hash)
• type Batched
• type BatchedBLAS
• type BatchedDevice
• type BinaryOp
• type BroadcastPattern
• func NewBroadcastPattern(leftAxes, rightAxes []byte) BroadcastPattern
• type CLDoer
• type CUDAADOp
• type CUDADoer
• type CloneErrorer
• type Cloner
• type CopierFrom
• type CopierTo
• type Device
• func (d Device) Alloc(extern External, size int64) (tensor.Memory, error)
• func (d Device) Free(extern External, mem tensor.Memory, sie uint) error
• func (d Device) IsGPU() bool
• func (d Device) String() string
• type DimSizer
• func ShapesToDimSizers(shapes []tensor.Shape) []DimSizer
• type Dtyper
• type Errer
• type ExecutionContext
• type ExprGraph
• func NewGraph(opts ...graphconopt) *ExprGraph
• func (g *ExprGraph) AddNode(n *Node) (retVal *Node)
• func (g *ExprGraph) AllNodes() Nodes
• func (g *ExprGraph) ByName(name string) (retVal Nodes)
• func (g *ExprGraph) Clone() interface{}
• func (g *ExprGraph) Constant(v Value) *Node
• func (g *ExprGraph) Edge(u, v int64) graph.Edge
• func (g *ExprGraph) Edges() graph.Edges
• func (g *ExprGraph) ExactSubgraphRoots(ns ...*Node) *ExprGraph
• func (g *ExprGraph) From(nodeid int64) graph.Nodes
• func (g *ExprGraph) Has(nodeid int64) bool
• func (g *ExprGraph) HasEdgeBetween(x, y int64) bool
• func (g *ExprGraph) HasEdgeFromTo(u, v int64) bool
• func (g *ExprGraph) Inputs() (retVal Nodes)
• func (g *ExprGraph) Node(id int64) graph.Node
• func (g *ExprGraph) Nodes() graph.Nodes
• func (g *ExprGraph) RemoveNode(node graph.Node)
• func (g *ExprGraph) Roots() (retVal Nodes)
• func (g *ExprGraph) SetEdge(e graph.Edge)
• func (g *ExprGraph) String() string
• func (g *ExprGraph) Subgraph(ns ...*Node) *ExprGraph
• func (g *ExprGraph) SubgraphRoots(ns ...*Node) *ExprGraph
• func (g *ExprGraph) To(nid int64) graph.Nodes
• func (g *ExprGraph) ToDot() string
• func (g *ExprGraph) UnbindAll()
• func (g *ExprGraph) UnbindAllNonInputs()
• type ExternMetadata
• func (m *ExternMetadata) Cleanup()
• func (m *ExternMetadata) DoWork() error
• func (m *ExternMetadata) Get(dev Device, size int64) (tensor.Memory, error)
• func (m *ExternMetadata) GetFromValue(dev Device, v Value) (tensor.Memory, error)
• func (m ExternMetadata) HasFunc(name string) bool
• func (m *ExternMetadata) Put(dev Device, mem tensor.Memory, size int64)
• func (m *ExternMetadata) PutValue(dev Device, v Value)
• func (m *ExternMetadata) Reset()
• func (m *ExternMetadata) Signal()
• func (m *ExternMetadata) Sync() chan struct{}
• func (m *ExternMetadata) Transfer(toDev, fromDev Device, v Value, synchronous bool) (retVal Value, err error)
• func (m *ExternMetadata) WorkAvailable() <-chan bool
• type External
• type ExternalOp
• func NewAddOp(a, b *Node, ctx ExecutionContext) *ExternalOp
• func NewExternalOp(op Op, ctx ExecutionContext, prealloc Value) *ExternalOp
• func NewHadamardProdOp(a, b *Node, ctx ExecutionContext) *ExternalOp
• func NewSubOp(a, b *Node, ctx ExecutionContext) *ExternalOp
• func (op *ExternalOp) DetermineDevice(inputs Nodes, output *Node) error
• func (op *ExternalOp) Do(vals ...Value) (Value, error)
• func (op *ExternalOp) String() string
• type F32
• func (v *F32) Data() interface{}
• func (v *F32) Dtype() tensor.Dtype
• func (v *F32) Format(s fmt.State, c rune)
• func (v *F32) MemSize() uintptr
• func (v *F32) Pointer() unsafe.Pointer
• func (v *F32) Shape() tensor.Shape
• func (v *F32) Size() int
• func (v *F32) Uintptr() uintptr
• type F64
• func (v *F64) Data() interface{}
• func (v *F64) Dtype() tensor.Dtype
• func (v *F64) Format(s fmt.State, c rune)
• func (v *F64) MemSize() uintptr
• func (v *F64) Pointer() unsafe.Pointer
• func (v *F64) Shape() tensor.Shape
• func (v *F64) Size() int
• func (v *F64) Uintptr() uintptr
• type I
• func (v *I) Data() interface{}
• func (v *I) Dtype() tensor.Dtype
• func (v *I) Format(s fmt.State, c rune)
• func (v *I) MemSize() uintptr
• func (v *I) Pointer() unsafe.Pointer
• func (v *I) Shape() tensor.Shape
• func (v *I) Size() int
• func (v *I) Uintptr() uintptr
• type I32
• func (v *I32) Data() interface{}
• func (v *I32) Dtype() tensor.Dtype
• func (v *I32) Format(s fmt.State, c rune)
• func (v *I32) MemSize() uintptr
• func (v *I32) Pointer() unsafe.Pointer
• func (v *I32) Shape() tensor.Shape
• func (v *I32) Size() int
• func (v *I32) Uintptr() uintptr
• type I64
• func (v *I64) Data() interface{}
• func (v *I64) Dtype() tensor.Dtype
• func (v *I64) Format(s fmt.State, c rune)
• func (v *I64) MemSize() uintptr
• func (v *I64) Pointer() unsafe.Pointer
• func (v *I64) Shape() tensor.Shape
• func (v *I64) Size() int
• func (v *I64) Uintptr() uintptr
• type IncrDoer
• type InitWFn
• func Gaussian(mean, stdev float64) InitWFn
• func GlorotN(gain float64) InitWFn
• func GlorotU(gain float64) InitWFn
• func HeN(gain float64) InitWFn
• func HeU(gain float64) InitWFn
• func Ones() InitWFn
• func RangedFrom(start int) InitWFn
• func Uniform(low, high float64) InitWFn
• func ValuesOf(val interface{}) InitWFn
• func Zeroes() InitWFn
• type Input
• type Mker
• type Momentum
• func NewMomentum(opts ...SolverOpt) *Momentum
• func (s *Momentum) Step(model []ValueGrad) (err error)
• type Namer
• type NoOpError
• type NoRetOp
• type Node
• func Abs(a *Node) (*Node, error)
• func Add(a, b *Node) (*Node, error)
• func ApplyOp(op Op, children ...*Node) (retVal *Node, err error)
• func ApplyOpWithName(op Op, name string, children ...*Node) (retVal *Node, err error)
• func At(a *Node, coords ...int) (retVal *Node, err error)
• func Auto(op func(a, b *Node, leftPattern, rightPattern []byte) (*Node, error), a, b *Node) (*Node, error)
• func BatchedMatMul(a, b *Node, transes ...bool) (retVal *Node, err error)
• func BinaryXent(output, target *Node) (retVal *Node, err error)
• func BinomialRandomNode(g *ExprGraph, dt tensor.Dtype, trials, prob float64, shape ...int) *Node
• func BroadcastAdd(a, b *Node, leftPattern, rightPattern []byte) (*Node, error)
• func BroadcastEq(a, b *Node, retSame bool, leftPattern, rightPattern []byte) (*Node, error)
• func BroadcastGt(a, b *Node, retSame bool, leftPattern, rightPattern []byte) (*Node, error)
• func BroadcastGte(a, b *Node, retSame bool, leftPattern, rightPattern []byte) (*Node, error)
• func BroadcastHadamardDiv(a, b *Node, leftPattern, rightPattern []byte) (*Node, error)
• func BroadcastHadamardProd(a, b *Node, leftPattern, rightPattern []byte) (*Node, error)
• func BroadcastLt(a, b *Node, retSame bool, leftPattern, rightPattern []byte) (*Node, error)
• func BroadcastLte(a, b *Node, retSame bool, leftPattern, rightPattern []byte) (*Node, error)
• func BroadcastNe(a, b *Node, retSame bool, leftPattern, rightPattern []byte) (*Node, error)
• func BroadcastPow(a, b *Node, leftPattern, rightPattern []byte) (*Node, error)
• func BroadcastSub(a, b *Node, leftPattern, rightPattern []byte) (*Node, error)
• func ByIndices(x *Node, indices *Node, axis int) (*Node, error)
• func Ceil(a *Node) (*Node, error)
• func Concat(axis int, ns ...*Node) (retVal *Node, err error)
• func Conv1d(in, filter *Node, kernel, pad, stride, dilation int) (*Node, error)
• func Conv2d(im, filter *Node, kernelShape tensor.Shape, pad, stride, dilation []int) (retVal *Node, err error)
• func ConvType(x *Node, from, to tensor.Dtype) (*Node, error)
• func Cos(a *Node) (*Node, error)
• func Cube(a *Node) (*Node, error)
• func DiagFlat(a *Node) (*Node, error)
• func Div(a, b *Node) (retVal *Node, err error)
• func Dropout(x *Node, dropProb float64) (retVal *Node, err error)
• func Eq(a, b *Node, retSame bool) (*Node, error)
• func Exp(a *Node) (*Node, error)
• func Expm1(a *Node) (*Node, error)
• func Floor(a *Node) (*Node, error)
• func GaussianRandomNode(g *ExprGraph, dt tensor.Dtype, mean, stdev float64, shape ...int) *Node
• func GlobalAveragePool2D(x *Node) (*Node, error)
• func Gt(a, b *Node, retSame bool) (*Node, error)
• func Gte(a, b *Node, retSame bool) (*Node, error)
• func HadamardDiv(a, b *Node) (*Node, error)
• func HadamardProd(a, b *Node) (*Node, error)
• func Im2Col(n *Node, kernel, pad, stride, dilation tensor.Shape) (retVal *Node, err error)
• func Inverse(a *Node) (*Node, error)
• func InverseSqrt(a *Node) (*Node, error)
• func KeepDims(a *Node, expandLeft bool, fn func(a *Node) (*Node, error)) (*Node, error)
• func LeakyRelu(x *Node, alpha float64) (*Node, error)
• func Log(a *Node) (*Node, error)
• func Log1p(a *Node) (*Node, error)
• func Log2(a *Node) (*Node, error)
• func LogSumExp(a *Node, axis int) (retVal *Node, err error)
• func Lt(a, b *Node, retSame bool) (*Node, error)
• func Lte(a, b *Node, retSame bool) (*Node, error)
• func Max(a *Node, along ...int) (retVal *Node, err error)
• func MaxPool1D(x *Node, kernel, pad, stride int) (*Node, error)
• func MaxPool2D(x *Node, kernel tensor.Shape, pad, stride []int) (*Node, error)
• func Mean(a *Node, along ...int) (retVal *Node, err error)
• func Mish(a *Node) (retVal *Node, err error)
• func Mul(a, b *Node) (retVal *Node, err error)
• func Must(n *Node, err error, opts ...NodeConsOpt) *Node
• func Ne(a, b *Node, retSame bool) (*Node, error)
• func Neg(a *Node) (*Node, error)
• func NewConstant(v interface{}, opts ...NodeConsOpt) *Node
• func NewMatrix(g *ExprGraph, t tensor.Dtype, opts ...NodeConsOpt) *Node
• func NewScalar(g *ExprGraph, t tensor.Dtype, opts ...NodeConsOpt) *Node
• func NewTensor(g *ExprGraph, t tensor.Dtype, dims int, opts ...NodeConsOpt) *Node
• func NewUniqueNode(opts ...NodeConsOpt) *Node
• func NewVector(g *ExprGraph, t tensor.Dtype, opts ...NodeConsOpt) *Node
• func NodeFromAny(g *ExprGraph, any interface{}, opts ...NodeConsOpt) *Node
• func Norm(a *Node, axis, p int) (retVal *Node, err error)
• func OneHotVector(id, classes int, t tensor.Dtype, opts ...NodeConsOpt) *Node
• func OuterProd(a, b *Node) (retVal *Node, err error)
• func Pow(a, b *Node) (*Node, error)
• func Ravel(n *Node) (retVal *Node, err error)
• func Read(n *Node, into *Value) (retVal *Node)
• func Rectify(x *Node) (retVal *Node, err error)
• func ReduceAdd(nodes Nodes, opts ...NodeConsOpt) (retVal *Node, err error)
• func ReduceMul(nodes Nodes, opts ...NodeConsOpt) (retVal *Node, err error)
• func Reshape(n *Node, to tensor.Shape) (retVal *Node, err error)
• func Set(a, b *Node) (retVal *Node)
• func Sigmoid(a *Node) (*Node, error)
• func Sign(a *Node) (*Node, error)
• func Sin(a *Node) (*Node, error)
• func SizeOf(axis int, x *Node) (retVal *Node, err error)
• func Slice(n *Node, slices ...tensor.Slice) (retVal *Node, err error)
• func SoftMax(x *Node, axis ...int) (*Node, error)
• func Softplus(a *Node) (*Node, error)
• func Sparsemax(x *Node, axes ...int) (*Node, error)
• func Sqrt(a *Node) (*Node, error)
• func Square(a *Node) (*Node, error)
• func Sub(a, b *Node) (*Node, error)
• func Sum(a *Node, along ...int) (retVal *Node, err error)
• func Tanh(a *Node) (*Node, error)
• func Tensordot(aAxes []int, bAxes []int, a, b *Node) (retVal *Node, err error)
• func Transpose(n *Node, axes ...int) (retVal *Node, err error)
• func UniformRandomNode(g *ExprGraph, dt tensor.Dtype, low, high float64, shape ...int) *Node
• func Upsample2D(x *Node, scale int) (*Node, error)
• func YOLOv3(input *Node, anchors []float32, masks []int, netSize, numClasses int, ignoreTresh float32, targets ...*Node) (*Node, error)
• func (n *Node) Clone() (retVal interface{})
• func (n *Node) CloneTo(g *ExprGraph) *Node
• func (n *Node) DataSize() int
• func (n *Node) Deriv() *Node
• func (n *Node) DerivOf() Nodes
• func (n *Node) Device() Device
• func (n *Node) Dims() int
• func (n *Node) Dtype() tensor.Dtype
• func (n *Node) Err() error
• func (n *Node) Grad() (Value, error)
• func (n *Node) GradOnDevice(dev Device, extern External) (retVal Value, allocOnExtern bool, err error)
• func (n *Node) Graph() *ExprGraph
• func (n *Node) Groups() encoding.Groups
• func (n *Node) Hashcode() uint32
• func (n *Node) ID() int64
• func (n *Node) IsColVec() bool
• func (n *Node) IsMatrix() bool
• func (n *Node) IsRowVec() bool
• func (n *Node) IsScalar() bool
• func (n *Node) IsVar() bool
• func (n *Node) IsVec() bool
• func (n *Node) IsVector() bool
• func (n *Node) Name() string
• func (n *Node) Node() *Node
• func (n *Node) Nodes() Nodes
• func (n *Node) Op() Op
• func (n *Node) RestrictedToDot(up, down int) string
• func (n *Node) Shape() tensor.Shape
• func (n *Node) Strides() []int
• func (n *Node) String() string
• func (n *Node) ToDot() string
• func (n *Node) Type() hm.Type
• func (n *Node) Value() Value
• func (n *Node) ValueOnDevice(dev Device, extern External) (retVal Value, allocOnExtern bool, err error)
• func (n *Node) WriteHash(h hash.Hash32)
• type NodeConsOpt
• func In(g *ExprGraph) NodeConsOpt
• func WithChildren(children Nodes) NodeConsOpt
• func WithGrad(any interface{}) NodeConsOpt
• func WithGroupName(name string) NodeConsOpt
• func WithInit(fn InitWFn) NodeConsOpt
• func WithName(name string) NodeConsOpt
• func WithOp(op Op) NodeConsOpt
• func WithShape(shp ...int) NodeConsOpt
• func WithType(t hm.Type) NodeConsOpt
• func WithValue(any interface{}) NodeConsOpt
• type NodeSet
• func NewNodeSet(a ...*Node) NodeSet
• func (set NodeSet) Add(i *Node) bool
• func (set NodeSet) Cardinality() int
• func (set *NodeSet) Clear()
• func (set NodeSet) Clone() NodeSet
• func (set NodeSet) Contains(i *Node) bool
• func (set NodeSet) ContainsAll(i ...*Node) bool
• func (set NodeSet) Difference(other NodeSet) NodeSet
• func (set NodeSet) Equal(other NodeSet) bool
• func (set NodeSet) Intersect(other NodeSet) NodeSet
• func (set NodeSet) IsSubset(other NodeSet) bool
• func (set NodeSet) IsSuperset(other NodeSet) bool
• func (set NodeSet) Iter() <-chan *Node
• func (set NodeSet) Remove(i *Node)
• func (set NodeSet) SymmetricDifference(other NodeSet) NodeSet
• func (set NodeSet) ToSlice() Nodes
• func (set NodeSet) Union(other NodeSet) NodeSet
• type Nodes
• func Backpropagate(outputs, gradOutputs, wrt Nodes) (retVal Nodes, err error)
• func Grad(cost *Node, WRTs ...*Node) (retVal Nodes, err error)
• func NodesFromInputs(xs ...Input) (Nodes, error)
• func Sort(g *ExprGraph) (sorted Nodes, err error)
• func Unconcat(a *Node, along int, n int) (Nodes, error)
• func UnstableSort(g *ExprGraph) (sorted Nodes, err error)
• func (ns Nodes) Add(n *Node) Nodes
• func (ns Nodes) AllSameGraph() bool
• func (ns Nodes) Contains(want *Node) bool
• func (ns Nodes) Difference(other Nodes) Nodes
• func (ns Nodes) Equals(other Nodes) bool
• func (ns Nodes) Err() error
• func (ns Nodes) Format(s fmt.State, c rune)
• func (ns Nodes) Intersect(other Nodes) Nodes
• func (ns Nodes) Len() int
• func (ns Nodes) Less(i, j int) bool
• func (ns Nodes) Node() *Node
• func (ns Nodes) Nodes() Nodes
• func (ns Nodes) Set() Nodes
• func (ns Nodes) Swap(i, j int)
• type Op
• type RMSPropSolver
• func NewRMSPropSolver(opts ...SolverOpt) *RMSPropSolver
• func (s *RMSPropSolver) Step(model []ValueGrad) (err error)
• type ReductionOp
• type Result
• func LiftResult(a Input, err error) Result
• type SDOp
• type Scalar
• type Solver
• type SolverOpt
• func WithBatchSize(batch float64) SolverOpt
• func WithBeta1(beta1 float64) SolverOpt
• func WithBeta2(beta2 float64) SolverOpt
• func WithClip(clip float64) SolverOpt
• func WithEps(eps float64) SolverOpt
• func WithL1Reg(l1reg float64) SolverOpt
• func WithL2Reg(l2reg float64) SolverOpt
• func WithLearnRate(eta float64) SolverOpt
• func WithMomentum(momentum float64) SolverOpt
• func WithRho(rho float64) SolverOpt
• type StandardEngine
• func (e StandardEngine) Transpose(a tensor.Tensor, expStrides []int) error
• type SymDiffError
• func (err SymDiffError) Error() string
• func (err SymDiffError) Grad() *Node
• func (err SymDiffError) Grads() map[*Node]Nodes
• func (err SymDiffError) Node() *Node
• func (err SymDiffError) Nodes() Nodes
• type Tensor
• type TensorType
• func (t TensorType) Apply(sub hm.Subs) hm.Substitutable
• func (t TensorType) Eq(other hm.Type) bool
• func (t TensorType) Format(state fmt.State, c rune)
• func (t TensorType) FreeTypeVar() hm.TypeVarSet
• func (t TensorType) Name() string
• func (t TensorType) Normalize(k, v hm.TypeVarSet) (hm.Type, error)
• func (t TensorType) String() string
• func (t TensorType) Types() hm.Types
• type Typer
• type U8
• func (v *U8) Data() interface{}
• func (v *U8) Dtype() tensor.Dtype
• func (v *U8) Format(s fmt.State, c rune)
• func (v *U8) MemSize() uintptr
• func (v *U8) Pointer() unsafe.Pointer
• func (v *U8) Shape() tensor.Shape
• func (v *U8) Size() int
• func (v *U8) Uintptr() uintptr
• type UnaryOp
• type UnsafeDoer
• type UsePreallocDoer
• type VM
• type VMOpt
• func BindDualValues(nodes ...*Node) VMOpt
• func ExecuteBwdOnly() VMOpt
• func ExecuteFwdOnly() VMOpt
• func LogBothDir() VMOpt
• func LogBwd() VMOpt
• func LogFwd() VMOpt
• func TraceExec() VMOpt
• func UseCudaFor(ops ...string) VMOpt
• func WithEngine(e tensor.Engine) VMOpt
• func WithInfWatch() VMOpt
• func WithLogger(logger *log.Logger) VMOpt
• func WithManualGradient() VMOpt
• func WithNaNWatch() VMOpt
• func WithPrecompiled(prog *program, locMap map[*Node]register) VMOpt
• func WithValueFmt(format string) VMOpt
• func WithWatchlist(list ...interface{}) VMOpt
• type Value
• func CloneValue(v Value) (Value, error)
• func Copy(dest, src Value) (Value, error)
• func ScalarAsTensor(v Value, dims int, e tensor.Engine) Value
• func ZeroValue(v Value) Value
• type ValueCloser
• type ValueEqualer
• type ValueGrad
• func NodesToValueGrads(in Nodes) (out []ValueGrad)
• type Valuer
• type VanillaSolver
• func NewVanillaSolver(opts ...SolverOpt) *VanillaSolver
• func (s *VanillaSolver) Step(model []ValueGrad) (err error)
• type ZeroValuer
• type Zeroer

Examples ¶

• BatchedMatMul
• BatchedMatMul (WithBackprop)
• Broadcast (Nils)
• BroadcastAdd
• BroadcastGte (CreatingTriangleMatrices)
• Concat
• DiagFlat
• SoftMax
• Tensordot (Vectors)
• Unconcat
• package (Autodiff)
• package (Basic)
• package (ConcurrentTraining)
• package (ErrorHandling)
• package (KeepDims)
• package (LinearRegression)
• package (NonConcurrentTraining)
• package (SymbolicDiff)

Package Files ¶

analysis.go api_gen.go batch.go bitmap.go blas.go broadcast.go collections.go compile.go concurrency.go const.go device.go differentiation.go doc.go dual.go engine.go equalities.go ermagerdmonards.go errors.go execution.go formatter.go gorgonia.go graph.go interfaces.go math.go math_nooptim.go mathutils_amd64.go nn.go node.go node_set.go noextern.go op.go op_by_indices.go op_infidel.go op_math.go op_math_noextern.go op_nn.go op_nondiff.go op_reduction.go op_softmax.go op_sparsemax.go op_tensor.go op_types.go op_upsample.go op_yolo.go operations.go operations_nondiff.go operatorLinAlg.go operatorLinAlg_const.go operatorPointwise_binary.go operatorPointwise_binary_const.go operatorPointwise_unary.go operatorPointwise_unary_const.go operatorPointwise_unary_gen.go opt.go perf.go regalloc.go release.go shape.go slice.go solvers.go stabilization.go templates.go type.go typeSystem.go utils.go values.go values_primitives.go values_utils.go vm.go vm_genera.go vm_genera_nocuda.go vm_tape.go vm_tape_nocuda.go walker.go weights.go

Constants ¶

const CUDA = false

CUDA indicates if this build is using CUDA

const DEBUG = false

DEBUG indicates if this build is in debug mode. It is not.

Variables ¶

var (

    // Float64 ...
    Float64 = tensor.Float64
    // Float32 ...
    Float32 = tensor.Float32
    // Int ...
    Int = tensor.Int
    // Int64 ...
    Int64 = tensor.Int64
    // Int32 ...
    Int32 = tensor.Int32
    // Byte ...
    Byte = tensor.Uint8
    // Bool ...
    Bool = tensor.Bool

    // Ptr is equivalent to interface{}. Ugh Ugh Ugh
    Ptr = tensor.UnsafePointer
)

func BatchNorm ¶ Uses

func BatchNorm(x, scale, bias *Node, momentum, epsilon float64) (retVal, γ, β *Node, op *BatchNormOp, err error)

BatchNorm applies a batchnormalization. This operator can be used in forward pass or for training. In an evaluation only, the "op" output can be discared. In training phase, γ, β can be discarded and the op should be used.

func BatchNorm1d ¶ Uses

func BatchNorm1d(x, scale, bias *Node, momentum, epsilon float64) (retVal, γ, β *Node, op *BatchNormOp, err error)

BatchNorm1d applies a batchnormalization to a matrix. This operator can be used in forward pass or for training. In an evaluation only, the "op" output can be discared. In training phase, γ, β can be discarded and the op should be used.

func Binomial32 ¶ Uses

func Binomial32(trials, prob float64, s ...int) []float32

Binomial32 returns a []float32 drawn from a binomial distribution given the trial and probability parameters.

func Binomial64 ¶ Uses

func Binomial64(trials, prob float64, s ...int) []float64

Binomial64 returns a []float64 drawn from a binomial distribution given the trial and probability parameters.

func Broadcast ¶ Uses

func Broadcast(a, b *Node, pattern BroadcastPattern) (*Node, *Node, error)

Broadcast apply the pattern to the input nodes and returns two nodes suitable for a binary operator. Broadcast works somewhat like Numpy's broadcast, except it's now exposed as a function.

g := NewGraph()
x := NewMatrix(g, Float64, WithShape(2, 3), WithName("x"))
y := NewMatrix(g, Float64, WithShape(2, 3), WithName("y"))
a, b, err := Broadcast(x, y, NewBroadcastPattern(nil, nil))
if err != nil {
    fmt.Printf("Error: %v\n", err)
    return
}
fmt.Printf("a == x %t; b == y %t", a == x, b == y)

a == x true; b == y true

func CheckOne ¶ Uses

func CheckOne(in Input) error

CheckOne checks whether an input is an error

func Compile ¶ Uses

func Compile(g *ExprGraph) (prog *program, locMap map[*Node]register, err error)

Compile takes a graph and outputs a program suitable for *tapeMachine to run

func CompileFunction ¶ Uses

func CompileFunction(g *ExprGraph, inputs, outputs Nodes) (prog *program, locMap map[*Node]register, err error)

CompileFunction takes a graph, subsets it based on the input and output nodes provided and outputs a program suitable for *tapeMachine to run. It is analogous to theano.Function(). If some input nodes are not used or is not reachable, this function will return an error

func DebugDerives ¶ Uses

func DebugDerives()

DebugDerives turns on the derivation debug option when printing a graph

func DimSizersToShapes ¶ Uses

func DimSizersToShapes(ds []DimSizer) ([]tensor.Shape, error)

DimSizersToShapes is a convenience function to convert a slice of DimSizer to a slice of tensor.Shape. It will return an error if any of them isn't a tensor.Shape

func DontDebugDerives ¶ Uses

func DontDebugDerives()

DontDebugDerives turns off derivation debug option when printing a graph. It is off by default

func Err ¶ Uses

func Err(e error) gErr

Err is a function that returns a gErr. It wraps errors with stack information. A gErr implements Result, as well as error. This way, the Err() method acts as an unwrapper.

func FmtNodeMap ¶ Uses

func FmtNodeMap(m interface{}) mapFmt

FmtNodeMap is a convenience function to print map[*Node]<T>

The fmt flag that makes it all nicely formatted is "-". Because a map consists of two types (key's type and val's type), and the Go fmt verb doesn't quite allow us to do something like "%ds", a hack is introduced to enable nicer printing of map[*Node]<T>

Here's the hack: The "#" flag is used to indicate if the map will use the Node's ID or Name when formatting the map.

%-v 	nodeName:%v
%-#v	nodeID:%v
%-d 	nodeName:%x
%-#d 	nodeID: %x
%-p 	nodeName:%p
%-#p	nodeID:%p

If the "-" flag is not found, then the formatter returns the default Go format for map[<T>]<T2>

func Gaussian32 ¶ Uses

func Gaussian32(mean, stdev float64, s ...int) []float32

Gaussian32 returns a []float32 drawn from a gaussian distribution as defined by the mean and stdev

func Gaussian64 ¶ Uses

func Gaussian64(mean, stdev float64, s ...int) []float64

Gaussian64 returns a []float64 drawn from a gaussian distribution as defined by the mean and stdev

func GlorotEtAlN32 ¶ Uses

func GlorotEtAlN32(gain float64, s ...int) []float32

GlorotEtAlN32 returns float32 weights sampled from a normal distribution using the methods specified in Glorot et. al (2010). See also: http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf

func GlorotEtAlN64 ¶ Uses

func GlorotEtAlN64(gain float64, s ...int) []float64

GlorotEtAlN64 returns float64 weights sampled from a normal distribution using the methods specified in Glorot et. al (2010). See also: http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf

func GlorotEtAlU32 ¶ Uses

func GlorotEtAlU32(gain float64, s ...int) []float32

GlorotEtAlU32 returns float32 weights sampled from a uniform distribution using the methods specified in Glorot et. al (2010). See also: http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf

For best results, use:

1.0 for gain for weights that will be used in linear and/or sigmoid units
math.Sqrt(2.0) for gain for weights that will be used in ReLU units
math.Sqrt(2.0 / (1+alpha*alpha)) for ReLU that are leaky with alpha

func GlorotEtAlU64 ¶ Uses

func GlorotEtAlU64(gain float64, s ...int) []float64

GlorotEtAlU64 returns float64 weights sampled from a uniform distribution using the methods specified in Glorot et. al (2010). See also: http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf

For best results, use:

1.0 for gain for weights that will be used in linear and/or sigmoid units
math.Sqrt(2.0) for gain for weights that will be used in ReLU units
math.Sqrt(2.0 / (1+alpha*alpha)) for ReLU that are leaky with alpha

func GraphCollisionStats ¶ Uses

func GraphCollisionStats() (int, int, int)

GraphCollisionStats returns the collisions in the graph only when built with the debug tag, otherwise it's a noop that returns 0

func HeEtAlN64 ¶ Uses

func HeEtAlN64(gain float64, s ...int) []float64

HeEtAlN64 returns float64 weights sampled from a normal distro, using the methods described in He et al (2015). The formula is:

randn(n) * sqrt(2/n)

See also https://arxiv.org/abs/1502.01852

For best results, use:

1.0 for gain for weights that will be used in linear and/or sigmoid units
math.Sqrt(2.0) for gain for weights that will be used in ReLU units
math.Sqrt(2.0 / (1+alpha*alpha)) for ReLU that are leaky with alpha

func HeEtAlU64 ¶ Uses

func HeEtAlU64(gain float64, s ...int) []float64

HeEtAlU64 returns float64 weights sampled from a uniform distro, using the methods described in He et al (2015). The formula is:

randn(n) * sqrt(2/n)

See also https://arxiv.org/abs/1502.01852

For best results, use:

1.0 for gain for weights that will be used in linear and/or sigmoid units
math.Sqrt(2.0) for gain for weights that will be used in ReLU units
math.Sqrt(2.0 / (1+alpha*alpha)) for ReLU that are leaky with alpha

func Let ¶ Uses

func Let(n *Node, be interface{}) error

Let binds a Value to a node that is a variable. A variable is represented as a *Node with no Op. It is equivalent to :

x = 2

func Lift1 ¶ Uses

func Lift1(fn func(a *Node) (*Node, error)) func(a Input) Result

Lift1 decorates a function with a precheck and post function lifting

func Lift1Axial ¶ Uses

func Lift1Axial(fn func(a *Node, axes ...int) (*Node, error)) func(a Input, axes ...int) Result

Lift1Axial decorates a function with a precheck and post function lifting

func Lift2 ¶ Uses

func Lift2(fn func(a, b *Node) (*Node, error)) func(a, b Input) Result

Lift2 decorates a function with a precheck and post function lifting

func Lift2Broadcast ¶ Uses

func Lift2Broadcast(fn func(a, b *Node, pat1, pat2 []byte) (*Node, error)) func(a, b Input, pat1, pat2 []byte) Result

Lift2Broadcast decorates a function with a precheck and post function lifting

func NewLispMachine ¶ Uses

func NewLispMachine(g *ExprGraph, opts ...VMOpt) *lispMachine

NewLispMachine creates a VM that executes the graph as it is traversed. Depending on the VMOpts passed in this VM is also capable of performing automatic differentiation.

func NewTapeMachine ¶ Uses

func NewTapeMachine(g *ExprGraph, opts ...VMOpt) *tapeMachine

NewTapeMachine creates a VM that compiles a graph into a prog.

func ReturnNode ¶ Uses

func ReturnNode(n *Node)

ReturnNode returns a node to the pool. It does not check that the *Node has been removed from the graph. USE WITH CAUTION.

func ReturnType ¶ Uses

func ReturnType(t hm.Type)

ReturnType ...

func S ¶ Uses

func S(start int, opt ...int) tensor.Slice

S creates a tensor.Slice. end is optional. It should be passed in as the first param of the optionals. step is optional. It should be passed in as the second param of the optionals.

Default end is start+1. Default step is 1, unless end == step+1, then it defaults to 0

func SetDerivOf ¶ Uses

func SetDerivOf(deriv, of *Node)

SetDerivOf is used to hack around the fundamental limitations of Gorgonia.

Specifically it is used to set a node as the derivative of another node, used in the cuDNN version of batch norm.

The cuDNN BatchNorm operation produces the derivatives for the scale and bias as a side effect of calculating the derivative of the input. Because Gorgonia's Ops are modelled as pure functions (and no tuples) this causes a bit of trouble. With the clever use of scratch space ops multireturn can be simulated. But this causes derivatives to not be set correctly.

func SetOptimizationLevel ¶ Uses

func SetOptimizationLevel(i int)

SetOptimizationLevel sets the fast math optimization level. By default, fast math is turned off, and this function is a no-op.

Use the `fastmath` build tag to use fast math

func TransformResult ¶ Uses

func TransformResult(ins ...Input) func(a Input, err error) Result

TransformResult is like LiftResult, but allows for custom data types that fulfil Mker

func TypeOf ¶ Uses

func TypeOf(v Value) hm.Type

TypeOf returns the Type of the value

func Uniform32 ¶ Uses

func Uniform32(low, high float64, s ...int) []float32

Uniform32 returns a []float64 drawn from a uniform distribution between [low, high) that is provided

func Uniform64 ¶ Uses

func Uniform64(low, high float64, s ...int) []float64

Uniform64 returns a []float64 drawn from a uniform distribution between [low, high) that is provided

func UnsafeLet ¶ Uses

func UnsafeLet(n *Node, be interface{}) error

UnsafeLet binds a Value to any node, not just a variable node. This means that you can use it to change any node's value at the runtime of the graph. UNSAFE!

Additional notes: if `be` is a tensor.Slice, and the node's op is a sliceOp or sliceIncrOp, the op's slice will be replaced with the new slice.

func Use ¶ Uses

func Use(b BLAS)

Use defines which BLAS implementation gorgonia should use. The default is Gonum's Native. These are the other options:

Use(blase.Implementation())
Use(cubone.Implementation())
Use(cgo.Implementation)

Note the differences in the brackets. The blase and cubone ones are functions.

func UseNonStable ¶ Uses

func UseNonStable()

UseNonStable turns off the stabilization functions when building graphs.

func UseStabilization ¶ Uses

func UseStabilization()

UseStabilization sets the global option to invoke stabilization functions when building the graph. Numerical stabilization is on by default

func ValueClose ¶ Uses

func ValueClose(a, b Value) bool

ValueClose checks whether two values are close to one another. It's predominantly used as an alternative equality test for floats

func ValueEq ¶ Uses

func ValueEq(a, b Value) bool

ValueEq is the equality function for values

func WalkGraph ¶ Uses

func WalkGraph(start *Node) <-chan *Node

WalkGraph walks a graph. It returns a channel of *Nodes, so be sure to consume the channel or there may be a deadlock

func WithGraphName ¶ Uses

func WithGraphName(name string) graphconopt

WithGraphName is a ExprGraph construction option that provides a name.

type ADOp ¶ Uses

type ADOp interface {
    Op

    DoDiff(ctx ExecutionContext, inputs Nodes, output *Node) error
}

An ADOp is an Op that supports automatic differentiation.

type AdaGradSolver ¶ Uses

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

AdaGradSolver is the solver that does adaptive gradient descent. Read the paper: http://jmlr.org/papers/v12/duchi11a.html

func NewAdaGradSolver ¶ Uses

func NewAdaGradSolver(opts ...SolverOpt) *AdaGradSolver

NewAdaGradSolver creates a new AdaGradSolver with sane-ish default values

func (*AdaGradSolver) Step ¶ Uses

func (s *AdaGradSolver) Step(model []ValueGrad) (err error)

Step steps through each node in the model and applies the Adaptive Gradient gradient descent algorithm on the value.

This function will error out if the nodes do not have an associated Grad value.

type AdamSolver ¶ Uses

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

AdamSolver is the Adaptive Moment Estimation solver (basically RMSProp on steroids). Paper: http://arxiv.org/abs/1412.6980

We overload the purpose of existing data structure of a *dualValue. However, instead of just holding a value and its derivative, the cache's *dualValues hold the Means of gradients (in .Value) and the variances of the gradients (in .d)

func NewAdamSolver ¶ Uses

func NewAdamSolver(opts ...SolverOpt) *AdamSolver

NewAdamSolver creates an Adam solver with these default values:

eta (learn rate)	  	: 0.001
eps (smoothing factor)		: 1e-8
beta1				: 0.9
beta2 				: 0.999
batch				: 1

func (*AdamSolver) Step ¶ Uses

func (s *AdamSolver) Step(model []ValueGrad) (err error)

Step steps through each node in the model and applies the Adaptive Moment Estimation gradient descent algorithm on the value.

This function will error out if the nodes do not have an associated Grad value.

type Arena ¶ Uses

type Arena interface {
    Get(dev Device, size int64) (tensor.Memory, error)       // Get returns a NoOpError when it cannot get a memory. Please allocate
    GetFromValue(dev Device, v Value) (tensor.Memory, error) // Gets a memory and copies the values into the memory and returns it.
    Put(dev Device, mem tensor.Memory, size int64)           // puts the memory back into the arena
    PutValue(dev Device, v Value)                            // puts the memory back into the arena

    // Transfers memory from device to device
    Transfer(toDev, fromDev Device, v Value, synchronous bool) (retVal Value, err error)
}

Arena is a representation of a pool of tensor.Memory

type AutoDiffError ¶ Uses

type AutoDiffError struct{}

AutoDiffError is an error which should be passed if the function is not differentiable. This is useful for Op implementations

func (AutoDiffError) Error ¶ Uses

func (err AutoDiffError) Error() string

type B ¶ Uses

type B bool

B represents a bool value.

func (*B) Data ¶ Uses

func (v *B) Data() interface{}

Data returns the original representation of the Value

func (*B) Dtype ¶ Uses

func (v *B) Dtype() tensor.Dtype

Dtype returns the Dtype of the value

func (*B) Format ¶ Uses

func (v *B) Format(s fmt.State, c rune)

Format implements fmt.Formatter

func (*B) MemSize ¶ Uses

func (v *B) MemSize() uintptr

MemSize satisfies the tensor.Memory interface

func (*B) Pointer ¶ Uses

func (v *B) Pointer() unsafe.Pointer

Pointer returns the pointer as an unsafe.Pointer. Satisfies the tensor.Memory interface

func (*B) Shape ¶ Uses

func (v *B) Shape() tensor.Shape

Shape returns a scalar shape for all scalar values

func (*B) Size ¶ Uses

func (v *B) Size() int

Size returns 0 for all scalar Values

func (*B) Uintptr ¶ Uses

func (v *B) Uintptr() uintptr

Uintptr satisfies the tensor.Memory interface

type BLAS ¶ Uses

type BLAS interface {
    blas.Float32
    blas.Float64
    blas.Complex64
    blas.Complex128
}

BLAS represents all the possible implementations of BLAS. The default is Gonum's Native

func WhichBLAS ¶ Uses

func WhichBLAS() BLAS

WhichBLAS returns the BLAS that gorgonia uses.

type BarzilaiBorweinSolver ¶ Uses

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

BarzilaiBorweinSolver / Barzilai-Borwein performs Gradient Descent in steepest descend direction Solves 0 = F(x), by

xᵢ₊₁ = xᵢ - eta * Grad(F)(xᵢ)

Where the learn rate eta is calculated by the Barzilai-Borwein method:

eta(xᵢ) = <(xᵢ - xᵢ₋₁), (Grad(F)(xᵢ) - Grad(F)(xᵢ₋₁))> /
                ∥(Grad(F)(xᵢ) - Grad(F)(xᵢ₋₁))∥²

The input learn rate is used for the first iteration.

TODO: Check out stochastic implementations, e.g. "Barzilai-Borwein Step Size for Stochastic Gradient Descent" https://arxiv.org/abs/1605.04131

func NewBarzilaiBorweinSolver ¶ Uses

func NewBarzilaiBorweinSolver(opts ...SolverOpt) *BarzilaiBorweinSolver

NewBarzilaiBorweinSolver creates a new Barzilai-Borwein solver withs some default values: the learn rate is set to 0.001 and the solver does not use clipping.

func (*BarzilaiBorweinSolver) Step ¶ Uses

func (s *BarzilaiBorweinSolver) Step(model []ValueGrad) (err error)

Step steps through each node in the model and applies the Barzilai-Borwein gradient descent algorithm on the value.

This function will error out if the nodes do not have an associated Grad value.

type BatchNormOp ¶ Uses

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

BatchNormOp is a batch normalization process as described by Ioffe and Szegedy (2015) - http://arxiv.org/abs/1502.03167

Normalization is done as:

γ(x - μ) / σ + β

γ is the scaling factor and β is the offset factor. These are created by BatchNorm()

func (*BatchNormOp) Arity ¶ Uses

func (op *BatchNormOp) Arity() int

Arity returns 1

func (*BatchNormOp) CallsExtern ¶ Uses

func (op *BatchNormOp) CallsExtern() bool

CallsExtern is false

func (*BatchNormOp) DiffWRT ¶ Uses

func (op *BatchNormOp) DiffWRT(inputs int) []bool

DiffWRT ...

func (*BatchNormOp) Do ¶ Uses

func (op *BatchNormOp) Do(values ...Value) (retVal Value, err error)

Do performs the batchnorm computation on the values

func (*BatchNormOp) DoDiff ¶ Uses

func (op *BatchNormOp) DoDiff(ctx ExecutionContext, inputs Nodes, output *Node) error

DoDiff does the gradient computation

func (*BatchNormOp) Hashcode ¶ Uses

func (op *BatchNormOp) Hashcode() uint32

Hashcode ...

func (*BatchNormOp) InferShape ¶ Uses

func (op *BatchNormOp) InferShape(ns ...DimSizer) (tensor.Shape, error)

InferShape from the input values

func (*BatchNormOp) OverwritesInput ¶ Uses

func (op *BatchNormOp) OverwritesInput() int

OverwritesInput is -1 (operator doesn't overwrite any input value)

func (*BatchNormOp) Reset ¶ Uses

func (op *BatchNormOp) Reset() error

Reset the operator by zeroing the internals scratch spaces

func (*BatchNormOp) ReturnsPtr ¶ Uses

func (op *BatchNormOp) ReturnsPtr() bool

ReturnsPtr is true

func (*BatchNormOp) SetTesting ¶ Uses

func (op *BatchNormOp) SetTesting()

SetTesting configure the op for testing mode

func (*BatchNormOp) SetTraining ¶ Uses

func (op *BatchNormOp) SetTraining()

SetTraining configure the op for training mode. A call to this function implicitly calls the Reset() method

func (*BatchNormOp) String ¶ Uses

func (op *BatchNormOp) String() string

func (*BatchNormOp) SymDiff ¶ Uses

func (op *BatchNormOp) SymDiff(inputs Nodes, output *Node, grad *Node) (retVal Nodes, err error)

SymDiff ...

func (*BatchNormOp) Type ¶ Uses

func (op *BatchNormOp) Type() hm.Type

Type ...

func (*BatchNormOp) UsePreallocDo ¶ Uses

func (op *BatchNormOp) UsePreallocDo(prealloc Value, inputs ...Value) (retVal Value, err error)

UsePreallocDo ...

func (*BatchNormOp) WriteHash ¶ Uses

func (op *BatchNormOp) WriteHash(h hash.Hash)

WriteHash ...

type Batched ¶ Uses

type Batched interface {
    WorkAvailable() <-chan struct{}
    DoWork()
}

Batched interface describes any object that can process batch work

type BatchedBLAS ¶ Uses

type BatchedBLAS interface {
    Batched
    BLAS
}

BatchedBLAS interface describes any object that can process BLAS work in batch

type BatchedDevice ¶ Uses

type BatchedDevice interface {
    Batched
    Retval() interface{}
    Errors() error
}

BatchedDevice is the superset of BatchedBLAS and the batched CUDA workflow.

type BinaryOp ¶ Uses

type BinaryOp interface {
    Op

    IsBinary() bool
}

A BinaryOp is an Op that takes only two inputs

type BroadcastPattern ¶ Uses

type BroadcastPattern byte

BroadcastPattern is actually a bit array. It's split into 2 nibbles - the left nibble represents the left operand, the right nibble represents the right operand:

xxxx|xxxx

The least significant bit of each nibble is elem 0. Concrete examples:

00000010 (0x02) = broadcast axis 1 of the right operand
00000001 (0x01) = broadcast axis 0 of the right operand
00000101 (0x09) = broadcast axis 0 AND axis 2 of the right operand
00010000 (0x10) = broadcast axis 0 of the left operand
00110000 (0x30) = broadcast axis 0 and axis 1 of the lef operand

You get the drill.

Do note that the current limitation of the BroadcastPattern allows only up to 4 dimensions per operand.

func NewBroadcastPattern ¶ Uses

func NewBroadcastPattern(leftAxes, rightAxes []byte) BroadcastPattern

NewBroadcastPattern is a helper function to create broadcast patterns

type CLDoer ¶ Uses

type CLDoer interface {
    CLDo(inputs ...Value) (Value, error)
}

CLDoer uses OpenCL to perform the Op. As of now, there are NO Ops that support OpenCL

type CUDAADOp ¶ Uses

type CUDAADOp interface {
    ADOp
    CUDADoDiff(extern External, dev Device, inputs Nodes, output *Node) error
}

A CUDAADOp operation have a specific method to run with CUDA

type CUDADoer ¶ Uses

type CUDADoer interface {
    CUDADo(extern External, dev Device, prealloc Value, inputs ...Value) (retVal Value, err error)
}

CUDADoer uses CUDA to perform the Op.

type CloneErrorer ¶ Uses

type CloneErrorer interface {
    Clone() (interface{}, error)
}

CloneErrorer represents any type that can clone itself and return an error if necessary

type Cloner ¶ Uses

type Cloner interface {
    Clone() interface{}
}

Cloner represents any type that can clone itself.

type CopierFrom ¶ Uses

type CopierFrom interface {
    CopyFrom(src interface{}) error
}

CopierFrom represents any type that can copy data from the source provided.

type CopierTo ¶ Uses

type CopierTo interface {
    CopyTo(dest interface{}) error
}

CopierTo represents any type that can copy data to the destination.

type Device ¶ Uses

type Device int

Device represents the device where the code will be executed on. In this build, all code will run on the CPU

const (
    // CPU the only device the graph will be executed on
    CPU Device = 0
)

func (Device) Alloc ¶ Uses

func (d Device) Alloc(extern External, size int64) (tensor.Memory, error)

Alloc allocates memory on the device. This is currently a NO-OP in this build

func (Device) Free ¶ Uses

func (d Device) Free(extern External, mem tensor.Memory, sie uint) error

Free frees the memory on the device. This is currently a NO-OP in this build

func (Device) IsGPU ¶ Uses

func (d Device) IsGPU() bool

IsGPU will always return false in this build

func (Device) String ¶ Uses

func (d Device) String() string

String implements fmt.Stringer and runtime.Stringer

type DimSizer ¶ Uses

type DimSizer interface {
    DimSize(int) (int, error)
}

DimSizer is any type (typically a tensor.Shape) that allows querying for a dimension size given an input dimension.

func ShapesToDimSizers ¶ Uses

func ShapesToDimSizers(shapes []tensor.Shape) []DimSizer

ShapesToDimSizers is a convenience function to convert a slice of tensor.Shape to a slice of DimSizer

type Dtyper ¶ Uses

type Dtyper interface {
    Dtype() tensor.Dtype
}

Dtyper represents any type (typically a Value) that knows its own Dtype

type Errer ¶ Uses

type Errer interface {
    Err() error
}

Errer is an interface that can return an error.

type ExecutionContext ¶ Uses

type ExecutionContext struct {
    External
    Device
}

ExecutionContext informs how an op should be executed

type ExprGraph ¶ Uses

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

ExprGraph is a data structure for a directed acyclic graph (of expressions). This structure is the main entry point for Gorgonia.

func NewGraph ¶ Uses

func NewGraph(opts ...graphconopt) *ExprGraph

NewGraph creates a new graph. Duh

func (*ExprGraph) AddNode ¶ Uses

func (g *ExprGraph) AddNode(n *Node) (retVal *Node)

AddNode adds n to the graph. It panics if the added node ID matches an existing node ID.

func (*ExprGraph) AllNodes ¶ Uses

func (g *ExprGraph) AllNodes() Nodes

AllNodes is like Nodes, but returns Nodes instead of []graph.Node. Nodes() has been reserved for the graph.Directed interface, so this one is named AllNodes instead

func (*ExprGraph) ByName ¶ Uses

func (g *ExprGraph) ByName(name string) (retVal Nodes)

ByName returns nodes that have the name provided. Bear in mind that the name that is compared to is the internal name, not the result of calling node.Name(). The reason for doing this is for ease of finding only names that are user-supplied, instead of autogenerated names

func (*ExprGraph) Clone ¶ Uses

func (g *ExprGraph) Clone() interface{}

Clone clones the graph. All nodes gets cloned, and their values are cloned as well.

func (*ExprGraph) Constant ¶ Uses

func (g *ExprGraph) Constant(v Value) *Node

Constant returns a constant that may be found in the graph. If no constant were found, a new one is created instead

func (*ExprGraph) Edge ¶ Uses

func (g *ExprGraph) Edge(u, v int64) graph.Edge

Edge returns the edge from u to v if such an edge exists and nil otherwise. The node v must be directly reachable from u as defined by the From method.

func (*ExprGraph) Edges ¶ Uses

func (g *ExprGraph) Edges() graph.Edges

Edges returns all the edges in the graph.

func (*ExprGraph) ExactSubgraphRoots ¶ Uses

func (g *ExprGraph) ExactSubgraphRoots(ns ...*Node) *ExprGraph

ExactSubgraphRoots creates a subgraph from the roots provided. The difference between SubgraphRoots and ExactSubgraphRoots is that ExactSubGraphRoots will not attempt to discover if any nodes are missing.

Given a function like the following:

z = x + y
set(x, -x.Grad) // setting the value of x to the negative of the gradient

When SubgraphRoots is used on z, the `-x.Grad` will be included. When using ExactSubgraphRoots, only `x` and `y` are included in the subgraph

func (*ExprGraph) From ¶ Uses

func (g *ExprGraph) From(nodeid int64) graph.Nodes

From returns all nodes in g that can be reached directly from n.

func (*ExprGraph) Has ¶ Uses

func (g *ExprGraph) Has(nodeid int64) bool

Has returns whether the node exists within the graph.

func (*ExprGraph) HasEdgeBetween ¶ Uses

func (g *ExprGraph) HasEdgeBetween(x, y int64) bool

HasEdgeBetween returns whether an edge exists between nodes x and y without considering direction.

func (*ExprGraph) HasEdgeFromTo ¶ Uses

func (g *ExprGraph) HasEdgeFromTo(u, v int64) bool

HasEdgeFromTo returns whether an edge exists in the graph from u to v.

func (*ExprGraph) Inputs ¶ Uses

func (g *ExprGraph) Inputs() (retVal Nodes)

Inputs returns a list of nodes which are inputs (that is to say, the user is required to set a value in it)

func (*ExprGraph) Node ¶ Uses

func (g *ExprGraph) Node(id int64) graph.Node

Node returns the node in the graph with the given ID.

func (*ExprGraph) Nodes ¶ Uses

func (g *ExprGraph) Nodes() graph.Nodes

Nodes returns all the nodes in the graph.

func (*ExprGraph) RemoveNode ¶ Uses

func (g *ExprGraph) RemoveNode(node graph.Node)

RemoveNode removes n from the graph, as well as any edges attached to it. If the node is not in the graph it is a no-op.

func (*ExprGraph) Roots ¶ Uses

func (g *ExprGraph) Roots() (retVal Nodes)

Roots returns a list of nodes that are not children of any other nodes

func (*ExprGraph) SetEdge ¶ Uses

func (g *ExprGraph) SetEdge(e graph.Edge)

SetEdge adds e, an edge from one node to another. If the nodes do not exist, they are added. It will panic if the IDs of the e.From and e.To are equal.

func (*ExprGraph) String ¶ Uses

func (g *ExprGraph) String() string

func (*ExprGraph) Subgraph ¶ Uses

func (g *ExprGraph) Subgraph(ns ...*Node) *ExprGraph

Subgraph subsets a graph. This function has overloaded meanings - If only one node is passed in, it assumes that the one node is the root, otherwise, it treats ns as the subset of nodes to be included in the subgraph

func (*ExprGraph) SubgraphRoots ¶ Uses

func (g *ExprGraph) SubgraphRoots(ns ...*Node) *ExprGraph

SubgraphRoots creates a subgraph, assuming the provided nodes are roots to the new subgraph.

func (*ExprGraph) To ¶ Uses

func (g *ExprGraph) To(nid int64) graph.Nodes

To returns all nodes in g that can reach directly to n.

func (*ExprGraph) ToDot ¶ Uses

func (g *ExprGraph) ToDot() string

ToDot generates the graph in graphviz format. The use of this is to generate for the entire graph which may have multiple trees with different roots TODO: This is getting unwieldy. Perhaps refactor out into a ToDot(...Opt)?

func (*ExprGraph) UnbindAll ¶ Uses

func (g *ExprGraph) UnbindAll()

UnbindAll unbinds all the values from the nodes

func (*ExprGraph) UnbindAllNonInputs ¶ Uses

func (g *ExprGraph) UnbindAllNonInputs()

UnbindAllNonInputs unbinds all the values from nodes that aren't input nodes

type ExternMetadata ¶ Uses

type ExternMetadata struct {
    tensor.Engine
    // contains filtered or unexported fields
}

ExternMetadata is used to hold metadata about external execution devices. In this build, it's an empty struct because the default build doesn't use external devices to execute the graph on

func (*ExternMetadata) Cleanup ¶ Uses

func (m *ExternMetadata) Cleanup()

Cleanup cleans up the ancillary allocations made during the calling of batched external device function.

The reason for this method is due to the fact that there is currently no way to free memory while the context is still running without causing some weirdness to the CUDA calls.

This is a No-op in this build

func (*ExternMetadata) DoWork ¶ Uses

func (m *ExternMetadata) DoWork() error

DoWork flushes any batched cgo calls. In this build it only flushes the batched BLAS calls.

func (*ExternMetadata) Get ¶ Uses

func (m *ExternMetadata) Get(dev Device, size int64) (tensor.Memory, error)

Get allocates a memory of the size. In this build it returns a NoOpError.

func (*ExternMetadata) GetFromValue ¶ Uses

func (m *ExternMetadata) GetFromValue(dev Device, v Value) (tensor.Memory, error)

GetFromValue allocates a memory of the size of v. In this build it returns a NoOpError, and v itself

func (ExternMetadata) HasFunc ¶ Uses

func (m ExternMetadata) HasFunc(name string) bool

HasFunc will always return false in this build

func (*ExternMetadata) Put ¶ Uses

func (m *ExternMetadata) Put(dev Device, mem tensor.Memory, size int64)

Put puts a previously allocated memory slab of the provided size back into the pool. Currently this is a No-op in this build.

func (*ExternMetadata) PutValue ¶ Uses

func (m *ExternMetadata) PutValue(dev Device, v Value)

PutValue puts a previously allocated value into the pool. In this build, it is a noop.

func (*ExternMetadata) Reset ¶ Uses

func (m *ExternMetadata) Reset()

Reset is a noop function for compatibility with the Cuda build

func (*ExternMetadata) Signal ¶ Uses

func (m *ExternMetadata) Signal()

Signal sends a signal down the workavailable channel, telling the VM to call the DoWork method. Signal is a synchronous method

func (*ExternMetadata) Sync ¶ Uses

func (m *ExternMetadata) Sync() chan struct{}

Sync returns the sync channel

func (*ExternMetadata) Transfer ¶ Uses

func (m *ExternMetadata) Transfer(toDev, fromDev Device, v Value, synchronous bool) (retVal Value, err error)

Transfer transfers a value from device to device. In this build, it's a noop, returning the input value, and a nil error

func (*ExternMetadata) WorkAvailable ¶ Uses

func (m *ExternMetadata) WorkAvailable() <-chan bool

WorkAvailable returns a channel of empty struct, which is used to signal to the VM when there is work available. The VM will then call the DoWork method.

type External ¶ Uses

type External interface {
    Arena
    Signal() // signals the machine to do work
    Sync() chan struct{}
}

External is a representation of an external device (cuda/cgo/openCL), conceptually modelled as a machine.

type ExternalOp ¶ Uses

type ExternalOp struct {
    Op
    ExecutionContext

    Prealloc  Value
    Incr      Value // is this a Incr? IncrDoers have higher precedence over PreallocDo
    UseUnsafe bool  // Is this an unsafe op? Lowest of all "special" Dos
}

ExternalOp is an op that contains an external context. This allows for ops to be run without needing a VM

func NewAddOp ¶ Uses

func NewAddOp(a, b *Node, ctx ExecutionContext) *ExternalOp

NewAddOp creates a new *ExternalOp that wraps an add op

func NewExternalOp ¶ Uses

func NewExternalOp(op Op, ctx ExecutionContext, prealloc Value) *ExternalOp

NewExternalOp creates a new *ExternalOp.

func NewHadamardProdOp ¶ Uses

func NewHadamardProdOp(a, b *Node, ctx ExecutionContext) *ExternalOp

NewHadamardProdOp creates a new *ExternalOp that wraps a mul op

func NewSubOp ¶ Uses

func NewSubOp(a, b *Node, ctx ExecutionContext) *ExternalOp

NewSubOp creates a new *ExternalOp that wraps a sub op

func (*ExternalOp) DetermineDevice ¶ Uses

func (op *ExternalOp) DetermineDevice(inputs Nodes, output *Node) error

DetermineDevice ...

func (*ExternalOp) Do ¶ Uses

func (op *ExternalOp) Do(vals ...Value) (Value, error)

Do performs the op,

func (*ExternalOp) String ¶ Uses

func (op *ExternalOp) String() string

type F32 ¶ Uses

type F32 float32

F32 represents a float32 value.

func (*F32) Data ¶ Uses

func (v *F32) Data() interface{}

Data returns the original representation of the Value

func (*F32) Dtype ¶ Uses

func (v *F32) Dtype() tensor.Dtype

Dtype returns the Dtype of the value

func (*F32) Format ¶ Uses

func (v *F32) Format(s fmt.State, c rune)

Format implements fmt.Formatter

func (*F32) MemSize ¶ Uses

func (v *F32) MemSize() uintptr

MemSize satisfies the tensor.Memory interface

func (*F32) Pointer ¶ Uses

func (v *F32) Pointer() unsafe.Pointer

Pointer returns the pointer as an unsafe.Pointer. Satisfies the tensor.Memory interface

func (*F32) Shape ¶ Uses

func (v *F32) Shape() tensor.Shape

Shape returns a scalar shape for all scalar values

func (*F32) Size ¶ Uses

func (v *F32) Size() int

Size returns 0 for all scalar Values

func (*F32) Uintptr ¶ Uses

func (v *F32) Uintptr() uintptr

Uintptr satisfies the tensor.Memory interface

type F64 ¶ Uses

type F64 float64

F64 represents a float64 value.

func (*F64) Data ¶ Uses

func (v *F64) Data() interface{}

Data returns the original representation of the Value

func (*F64) Dtype ¶ Uses

func (v *F64) Dtype() tensor.Dtype

Dtype returns the Dtype of the value

func (*F64) Format ¶ Uses

func (v *F64) Format(s fmt.State, c rune)

Format implements fmt.Formatter

func (*F64) MemSize ¶ Uses

func (v *F64) MemSize() uintptr

MemSize satisfies the tensor.Memory interface

func (*F64) Pointer ¶ Uses

func (v *F64) Pointer() unsafe.Pointer

Pointer returns the pointer as an unsafe.Pointer. Satisfies the tensor.Memory interface

func (*F64) Shape ¶ Uses

func (v *F64) Shape() tensor.Shape

Shape returns a scalar shape for all scalar values

func (*F64) Size ¶ Uses

func (v *F64) Size() int

Size returns 0 for all scalar Values

func (*F64) Uintptr ¶ Uses

func (v *F64) Uintptr() uintptr

Uintptr satisfies the tensor.Memory interface

type I ¶ Uses

type I int

I represents a int value.

func (*I) Data ¶ Uses

func (v *I) Data() interface{}

Data returns the original representation of the Value

func (*I) Dtype ¶ Uses

func (v *I) Dtype() tensor.Dtype

Dtype returns the Dtype of the value

func (*I) Format ¶ Uses

func (v *I) Format(s fmt.State, c rune)

Format implements fmt.Formatter

func (*I) MemSize ¶ Uses

func (v *I) MemSize() uintptr

MemSize satisfies the tensor.Memory interface

func (*I) Pointer ¶ Uses

func (v *I) Pointer() unsafe.Pointer

Pointer returns the pointer as an unsafe.Pointer. Satisfies the tensor.Memory interface

func (*I) Shape ¶ Uses

func (v *I) Shape() tensor.Shape

Shape returns a scalar shape for all scalar values

func (*I) Size ¶ Uses

func (v *I) Size() int

Size returns 0 for all scalar Values

func (*I) Uintptr ¶ Uses

func (v *I) Uintptr() uintptr

Uintptr satisfies the tensor.Memory interface

type I32 ¶ Uses

type I32 int32

I32 represents a int32 value.

func (*I32) Data ¶ Uses

func (v *I32) Data() interface{}

Data returns the original representation of the Value

func (*I32) Dtype ¶ Uses

func (v *I32) Dtype() tensor.Dtype

Dtype returns the Dtype of the value

func (*I32) Format ¶ Uses

func (v *I32) Format(s fmt.State, c rune)

Format implements fmt.Formatter

func (*I32) MemSize ¶ Uses

func (v *I32) MemSize() uintptr

MemSize satisfies the tensor.Memory interface

func (*I32) Pointer ¶ Uses

func (v *I32) Pointer() unsafe.Pointer

Pointer returns the pointer as an unsafe.Pointer. Satisfies the tensor.Memory interface

func (*I32) Shape ¶ Uses

func (v *I32) Shape() tensor.Shape

Shape returns a scalar shape for all scalar values

func (*I32) Size ¶ Uses

func (v *I32) Size() int

Size returns 0 for all scalar Values

func (*I32) Uintptr ¶ Uses

func (v *I32) Uintptr() uintptr

Uintptr satisfies the tensor.Memory interface

type I64 ¶ Uses

type I64 int64

I64 represents a int64 value.

func (*I64) Data ¶ Uses

func (v *I64) Data() interface{}

Data returns the original representation of the Value

func (*I64) Dtype ¶ Uses

func (v *I64) Dtype() tensor.Dtype

Dtype returns the Dtype of the value

func (*I64) Format ¶ Uses

func (v *I64) Format(s fmt.State, c rune)

Format implements fmt.Formatter

func (*I64) MemSize ¶ Uses

func (v *I64) MemSize() uintptr

MemSize satisfies the tensor.Memory interface

func (*I64) Pointer ¶ Uses

func (v *I64) Pointer() unsafe.Pointer

Pointer returns the pointer as an unsafe.Pointer. Satisfies the tensor.Memory interface

func (*I64) Shape ¶ Uses

func (v *I64) Shape() tensor.Shape

Shape returns a scalar shape for all scalar values

func (*I64) Size ¶ Uses

func (v *I64) Size() int

Size returns 0 for all scalar Values

func (*I64) Uintptr ¶ Uses

func (v *I64) Uintptr() uintptr

Uintptr satisfies the tensor.Memory interface

type IncrDoer ¶ Uses

type IncrDoer interface {
    IncrDo(toIncr Value, inputs ...Value) error
}

IncrDoer increments the toIncr with the result of doing

type InitWFn ¶ Uses

type InitWFn func(dt tensor.Dtype, s ...int) interface{}

InitWFn is a type of helper function to help initialize weights vector/matrices. It generates the backing required for the tensors.

It's typically used in closures

func Gaussian ¶ Uses

func Gaussian(mean, stdev float64) InitWFn

Gaussian creates a InitWFn with the specified parameters. Example Usage:

w := NewMatrix(g, Float64, WithName("w"), WithShape(2,2), WithInit(Gaussian(0, 1)))

This will create a backing slice of []float64, with the length of 4, and its values are drawn from a gaussian distro

func GlorotN ¶ Uses

func GlorotN(gain float64) InitWFn

GlorotN creates a InitWFn that populates a Value with weights normally sampled using Glorot et al.'s algorithm

func GlorotU ¶ Uses

func GlorotU(gain float64) InitWFn

GlorotU creates a InitWFn that populates a Value with weights uniformly sampled using Glorot et al.'s algorithm

func HeN ¶ Uses

func HeN(gain float64) InitWFn

func HeU ¶ Uses

func HeU(gain float64) InitWFn

func Ones ¶ Uses

func Ones() InitWFn

Ones creates an InitWfn that populates a Value with ones. See Zeroes() for more explanation.

func RangedFrom ¶ Uses

func RangedFrom(start int) InitWFn

RangedFrom creates an InitWFn that populates a Value starting with the provided start, increamenting the number for each element in the value by 1

func Uniform ¶ Uses

func Uniform(low, high float64) InitWFn

Uniform creates a InitWFn with the specified parameters. Example Usage:

w := NewMatrix(g, Float64, WithName("w"), WithShape(2,2), WithInit(Uniform(-1, 1)))

This will create a backing slice of []float64, with the length of 4, and its values are drawn from a uniform distro

func ValuesOf ¶ Uses

func ValuesOf(val interface{}) InitWFn

ValuesOf creates an InitWrn that populates a value with val. This function will cause a panic if val's type is incompatible with the values type.

func Zeroes ¶ Uses

func Zeroes() InitWFn

Zeroes creates an InitWfn that populates a Value with... zeroes. I don't know what you expected.

type Input ¶ Uses

type Input interface {
    Node() *Node
    Nodes() Nodes
}

Input is something that can produce both a *Node and Nodes. Returning nil is OK.

type Mker ¶ Uses

type Mker interface {
    Mk(...Input) Input
}

Mker is an interface of any Input that can make a new version of itself

type Momentum ¶ Uses

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

Momentum is the stochastic gradient descent optimizer with momentum item.

func NewMomentum ¶ Uses

func NewMomentum(opts ...SolverOpt) *Momentum

NewMomentum creates a new Momentum with sane-ish default values

func (*Momentum) Step ¶ Uses

func (s *Momentum) Step(model []ValueGrad) (err error)

Step steps through each node in the model and applies the Momentum stochastic gradient descent algorithm on the value.

This function will error out if the nodes do not have an associated Grad value.

type Namer ¶ Uses

type Namer interface {
    Name() string
}

Namer is anything that has a name

type NoOpError ¶ Uses

type NoOpError interface {
    NoOp() bool
}

NoOpError is an error returned when an operation does nothing.

type NoRetOp ¶ Uses

type NoRetOp interface {
    Op

    ReturnsNothing() bool
}

A NoRetOp is an Op that reads a value, but does not return any value. It's a representation of a not-pure function

type Node ¶ Uses

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

A Node is a node in the computation graph

func Abs ¶ Uses

func Abs(a *Node) (*Node, error)

Abs performs a pointwise abs.

func Add ¶ Uses

func Add(a, b *Node) (*Node, error)

Add performs a pointwise add operation.

func ApplyOp ¶ Uses

func ApplyOp(op Op, children ...*Node) (retVal *Node, err error)

ApplyOp is the generic function application - for when no specialization is required

func ApplyOpWithName ¶ Uses

func ApplyOpWithName(op Op, name string, children ...*Node) (retVal *Node, err error)

ApplyOpWithName applies the op, and then gives the node the given name

func At ¶ Uses

func At(a *Node, coords ...int) (retVal *Node, err error)

At is a symbolic operation for getting a value at the provided coordinates. If the input is a scalar, all the coordinates MUST be 0, or else an error will be returned.

func Auto ¶ Uses

func Auto(op func(a, b *Node, leftPattern, rightPattern []byte) (*Node, error), a, b *Node) (*Node, error)

Auto automatically calculates the padding for the given operations, for example:

gorgonia.Auto(gorgonia.BroadcastHadamardProd, a, b)

func BatchedMatMul ¶ Uses

func BatchedMatMul(a, b *Node, transes ...bool) (retVal *Node, err error)

BatchedMatMul returns a node representing the batched mat mul operation.

A list of transpose options are allowed. The

g := NewGraph()
a := NewTensor(g, Float64, 3, WithShape(2, 2, 3), WithInit(RangedFrom(1)), WithName("a"))
b := NewTensor(g, Float64, 3, WithShape(2, 3, 2), WithInit(RangedFrom(13)), WithName("b"))
c, err := BatchedMatMul(a, b)
if err != nil {
    log.Fatal(err)
}
x := NewTensor(g, Float64, 4, WithShape(3, 2, 2, 3), WithInit(RangedFrom(1)), WithName("x"))
y := NewTensor(g, Float64, 4, WithShape(3, 2, 3, 2), WithInit(RangedFrom(37)), WithName("y"))
z, err := BatchedMatMul(x, y)
if err != nil {
    log.Fatal(err)
}

m := NewTapeMachine(g)
if err := m.RunAll(); err != nil {
    log.Fatal(err)
}

fmt.Printf("a: %v\n%v\n", a.Value().Shape(), a.Value().Data())
fmt.Printf("b: %v\n%v\n", b.Value().Shape(), b.Value().Data())
fmt.Printf("c: %v\n%v\n", c.Value().Shape(), c.Value().Data())
fmt.Printf("x: %v\n%v\n", x.Value().Shape(), x.Value().Data())
fmt.Printf("y: %v\n%v\n", y.Value().Shape(), y.Value().Data())
fmt.Printf("z: %v\n%v\n", z.Value().Shape(), z.Value().Data())

a: (2, 2, 3)
[1 2 3 4 5 6 7 8 9 10 11 12]
b: (2, 3, 2)
[13 14 15 16 17 18 19 20 21 22 23 24]
c: (2, 2, 2)
[94 100 229 244 508 532 697 730]
x: (3, 2, 2, 3)
[1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36]
y: (3, 2, 3, 2)
[37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72]
z: (3, 2, 2, 2)
[238 244 589 604 1084 1108 1489 1522 2146 2188 2605 2656 3424 3484 3937 4006 4918 4996 5485 5572 6628 6724 7249 7354]

g := NewGraph()
a := NewTensor(g, Float64, 4, WithShape(2, 4, 3, 9), WithInit(RangedFrom(1)), WithName("a"))
b := NewTensor(g, Float64, 4, WithShape(2, 4, 3, 9), WithInit(RangedFrom(13)), WithName("b"))
c, err := BatchedMatMul(a, b, false, true)
if err != nil {
    log.Fatal(err)
}
s, err := Sum(c)
if err != nil {
    log.Fatal(err)
}
grads, err := Grad(s, a, b)
if err != nil {
    log.Fatal(err)
}

m := NewTapeMachine(g)
if err := m.RunAll(); err != nil {
    log.Fatal(err)
}

fmt.Printf("a: %v\n%v\n", a.Value().Shape(), a.Value().Data())
fmt.Printf("b: %v\n%v\n", b.Value().Shape(), b.Value().Data())
fmt.Printf("c: %v\n%v\n", c.Value().Shape(), c.Value().Data())
fmt.Printf("grads[0]:%v\n%v\n", grads[0].Shape(), grads[0].Value().Data())

a: (2, 4, 3, 9)
[1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216]
b: (2, 4, 3, 9)
[13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228]
c: (2, 4, 3, 3)
[825 1230 1635 2202 3336 4470 3579 5442 7305 12732 15324 17916 16296 19617 22938 19860 23910 27960 37761 42540 47319 43512 49020 54528 49263 55500 61737 75912 82878 89844 83850 91545 99240 91788 100212 108636 127185 136338 145491 137310 147192 157074 147435 158046 168657 191580 202920 214260 203892 215961 228030 216204 229002 241800 269097 282624 296151 283596 297852 312108 298095 313080 328065 359736 375450 391164 376422 392865 409308 393108 410280 427452]
grads[0]:(2, 4, 3, 9)
[66 69 72 75 78 81 84 87 90 66 69 72 75 78 81 84 87 90 66 69 72 75 78 81 84 87 90 147 150 153 156 159 162 165 168 171 147 150 153 156 159 162 165 168 171 147 150 153 156 159 162 165 168 171 228 231 234 237 240 243 246 249 252 228 231 234 237 240 243 246 249 252 228 231 234 237 240 243 246 249 252 309 312 315 318 321 324 327 330 333 309 312 315 318 321 324 327 330 333 309 312 315 318 321 324 327 330 333 390 393 396 399 402 405 408 411 414 390 393 396 399 402 405 408 411 414 390 393 396 399 402 405 408 411 414 471 474 477 480 483 486 489 492 495 471 474 477 480 483 486 489 492 495 471 474 477 480 483 486 489 492 495 552 555 558 561 564 567 570 573 576 552 555 558 561 564 567 570 573 576 552 555 558 561 564 567 570 573 576 633 636 639 642 645 648 651 654 657 633 636 639 642 645 648 651 654 657 633 636 639 642 645 648 651 654 657]

func BinaryXent ¶ Uses

func BinaryXent(output, target *Node) (retVal *Node, err error)

BinaryXent is a convenience function for doing binary crossentropy stuff. The formula is as below:

-(y * logprob) +  (1-y)(1-logprob)

func BinomialRandomNode ¶ Uses

func BinomialRandomNode(g *ExprGraph, dt tensor.Dtype, trials, prob float64, shape ...int) *Node

BinomialRandomNode creates an input node that has a random op so that everytime the node is passed, random values will be plucked from a binomial distribution with the mean and stdev provided. The type of the node depends on the shape passed in. To get a scalar value at run time, don't pass in any shapes

Whilst technically the number of trials of a binomal distribution should be a discrete value (you can't have half a trial), to keep with API uniformity, trials is passed in as a float64, but will be truncated to an int at runtime.

func BroadcastAdd ¶ Uses

func BroadcastAdd(a, b *Node, leftPattern, rightPattern []byte) (*Node, error)

Add performs a add. The operation is precomposed with a broadcast such that the shapes matches before operations commence.

g := NewGraph()
a := NewVector(g, tensor.Float64, WithShape(2), WithName("a"), WithValue(tensor.New(tensor.WithBacking([]float64{100, 100}))))
b := NewMatrix(g, tensor.Float64, WithShape(2, 2), WithName("b"), WithValue(tensor.New(tensor.WithShape(2, 2), tensor.WithBacking([]float64{1, 1, 2, 2}))))

fmt.Printf("a = %v\nb =\n%v\n", a.Value(), b.Value())

_, err := Add(a, b)
fmt.Printf("a + b yields an error: %v\n\n", err)

// Note here the broadcasting of a is on the first axis, not the zeroth axis. Simply put, assume that it's already a (2,1) matrix.
ab, err := BroadcastAdd(a, b, []byte{1}, nil)
if err != nil {
    fmt.Printf("uh oh, something went wrong: %v\n", err)
}

ba, err := BroadcastAdd(b, a, nil, []byte{1})
if err != nil {
    fmt.Printf("uh oh, something went wrong: %v\n", err)
}

// Now, let's run the program
machine := NewTapeMachine(g)
defer machine.Close()
if err = machine.RunAll(); err != nil {
    log.Fatal(err)
}

fmt.Printf("a +⃗ b =\n%v\n", ab.Value())
fmt.Printf("b +⃗ a =\n%v", ba.Value())

a = [100  100]
b =
⎡1  1⎤
⎣2  2⎦

a + b yields an error: Failed to infer shape. Op: + false: Shape mismatch: (2) and (2, 2)

a +⃗ b =
⎡101  101⎤
⎣102  102⎦

b +⃗ a =
⎡101  101⎤
⎣102  102⎦

func BroadcastEq ¶ Uses

func BroadcastEq(a, b *Node, retSame bool, leftPattern, rightPattern []byte) (*Node, error)

Eq performs a eq. The operation is precomposed with a broadcast such that the shapes matches before operations commence.

func BroadcastGt ¶ Uses

func BroadcastGt(a, b *Node, retSame bool, leftPattern, rightPattern []byte) (*Node, error)

Gt performs a gt. The operation is precomposed with a broadcast such that the shapes matches before operations commence.

func BroadcastGte ¶ Uses

func BroadcastGte(a, b *Node, retSame bool, leftPattern, rightPattern []byte) (*Node, error)

Gte performs a gte. The operation is precomposed with a broadcast such that the shapes matches before operations commence.

// Broadcasting is useful. We can create triangular dense matrices simply

g := NewGraph()
a := NewMatrix(g, tensor.Float64, WithShape(3, 1), WithName("a"), WithInit(RangedFrom(0)))
b := NewMatrix(g, tensor.Float64, WithShape(1, 4), WithName("b"), WithInit(RangedFrom(0)))
tl, err := BroadcastGte(a, b, true, []byte{1}, []byte{0})
if err != nil {
    log.Fatalf("uh oh. Something went wrong %v", err)
}

tu, err := BroadcastLt(a, b, true, []byte{1}, []byte{0})
if err != nil {
    log.Fatalf("uh oh. Something went wrong %v", err)
}

m := NewTapeMachine(g)

// PEDAGOGICAL:
// Uncomment the following code if you want to see what happens behind the scenes
// m.Close()
// logger := log.New(os.Stderr, "",0)
// m = NewTapeMachine(g, WithLogger(logger), WithWatchlist())

defer m.Close()
if err = m.RunAll(); err != nil {
    log.Fatal(err)
}

fmt.Printf("triangular, lower:\n%v\n", tl.Value())
fmt.Printf("triangular, upper:\n%v\n", tu.Value())

triangular, lower:
⎡1  0  0  0⎤
⎢1  1  0  0⎥
⎣1  1  1  0⎦

triangular, upper:
⎡0  1  1  1⎤
⎢0  0  1  1⎥
⎣0  0  0  1⎦

func BroadcastHadamardDiv ¶ Uses

func BroadcastHadamardDiv(a, b *Node, leftPattern, rightPattern []byte) (*Node, error)

HadamardDiv performs a hadamarddiv. The operation is precomposed with a broadcast such that the shapes matches before operations commence.

func BroadcastHadamardProd ¶ Uses

func BroadcastHadamardProd(a, b *Node, leftPattern, rightPattern []byte) (*Node, error)

HadamardProd performs a hadamardprod. The operation is precomposed with a broadcast such that the shapes matches before operations commence.

func BroadcastLt ¶ Uses

func BroadcastLt(a, b *Node, retSame bool, leftPattern, rightPattern []byte) (*Node, error)

Lt performs a lt. The operation is precomposed with a broadcast such that the shapes matches before operations commence.

func BroadcastLte ¶ Uses

func BroadcastLte(a, b *Node, retSame bool, leftPattern, rightPattern []byte) (*Node, error)

Lte performs a lte. The operation is precomposed with a broadcast such that the shapes matches before operations commence.

func BroadcastNe ¶ Uses

func BroadcastNe(a, b *Node, retSame bool, leftPattern, rightPattern []byte) (*Node, error)

Ne performs a ne. The operation is precomposed with a broadcast such that the shapes matches before operations commence.

func BroadcastPow ¶ Uses

func BroadcastPow(a, b *Node, leftPattern, rightPattern []byte) (*Node, error)

Pow performs a pow. The operation is precomposed with a broadcast such that the shapes matches before operations commence.

func BroadcastSub ¶ Uses

func BroadcastSub(a, b *Node, leftPattern, rightPattern []byte) (*Node, error)

Sub performs a sub. The operation is precomposed with a broadcast such that the shapes matches before operations commence.

func ByIndices ¶ Uses

func ByIndices(x *Node, indices *Node, axis int) (*Node, error)

ByIndices is an operation that takes the indices as input and return the selected values from those indices. The default axis in 0

func Ceil ¶ Uses

func Ceil(a *Node) (*Node, error)

Ceil performs a pointwise ceil.

func Concat ¶ Uses

func Concat(axis int, ns ...*Node) (retVal *Node, err error)

Concat performs a concatenate on the provided axis and inputs.

g := NewGraph()
x := NewTensor(g, Float64, 4, WithShape(2, 3, 4, 5), WithInit(RangedFrom(0)), WithName("x"))
y := NewTensor(g, Float64, 4, WithShape(2, 3, 4, 5), WithInit(RangedFrom(120)), WithName("y"))

z, err := Concat(2, x, y)
if err != nil {
    panic(err)
}

m := NewTapeMachine(g)
if err := m.RunAll(); err != nil {
    panic(err)
}
tmp := fmt.Sprintf("z %v\n%v", z.Value().Shape(), z.Value())
fmt.Println(strings.Replace(tmp, "\n\n", "\n", -1)) // this is because

z (2, 3, 8, 5)
⎡  0    1    2    3    4⎤
⎢  5    6    7    8    9⎥
⎢ 10   11   12   13   14⎥
⎢ 15   16   17   18   19⎥
⎢120  121  122  123  124⎥
⎢125  126  127  128  129⎥
⎢130  131  132  133  134⎥
⎣135  136  137  138  139⎦

⎡ 20   21   22   23   24⎤
⎢ 25   26   27   28   29⎥
⎢ 30   31   32   33   34⎥
⎢ 35   36   37   38   39⎥
⎢140  141  142  143  144⎥
⎢145  146  147  148  149⎥
⎢150  151  152  153  154⎥
⎣155  156  157  158  159⎦

⎡ 40   41   42   43   44⎤
⎢ 45   46   47   48   49⎥
⎢ 50   51   52   53   54⎥
⎢ 55   56   57   58   59⎥
⎢160  161  162  163  164⎥
⎢165  166  167  168  169⎥
⎢170  171  172  173  174⎥
⎣175  176  177  178  179⎦

⎡ 60   61   62   63   64⎤
⎢ 65   66   67   68   69⎥
⎢ 70   71   72   73   74⎥
⎢ 75   76   77   78   79⎥
⎢180  181  182  183  184⎥
⎢185  186  187  188  189⎥
⎢190  191  192  193  194⎥
⎣195  196  197  198  199⎦

⎡ 80   81   82   83   84⎤
⎢ 85   86   87   88   89⎥
⎢ 90   91   92   93   94⎥
⎢ 95   96   97   98   99⎥
⎢200  201  202  203  204⎥
⎢205  206  207  208  209⎥
⎢210  211  212  213  214⎥
⎣215  216  217  218  219⎦

⎡100  101  102  103  104⎤
⎢105  106  107  108  109⎥
⎢110  111  112  113  114⎥
⎢115  116  117  118  119⎥
⎢220  221  222  223  224⎥
⎢225  226  227  228  229⎥
⎢230  231  232  233  234⎥
⎣235  236  237  238  239⎦

func Conv1d ¶ Uses

func Conv1d(in, filter *Node, kernel, pad, stride, dilation int) (*Node, error)

Conv1d is a 1D convlution. It relies on Conv2D

func Conv2d ¶ Uses

func Conv2d(im, filter *Node, kernelShape tensor.Shape, pad, stride, dilation []int) (retVal *Node, err error)

Conv2d is a simple 2D convolution, to be used for CPU computation only. If CuDNN is used, use the CUDAConv2D function. These are the properties the inputs must fulfil:

- im: must have 4D shape. Expected format is BCHW (batch, channels, height, width) - filter: must have 4D shape: (batch, kernel, height, width) - kernelShape: shape of the filter kernel - pad: len(pad) == 2, defaults to []int{0, 0} if nil is passed - stride: len(stride) == 2, example: []int{1, 1} - dilation: len(dilation) == 2, defaults to []int{1, 1} if nil is passed

func ConvType ¶ Uses

func ConvType(x *Node, from, to tensor.Dtype) (*Node, error)

ConvType converts the type of the x Node from one type to other

func Cos ¶ Uses

func Cos(a *Node) (*Node, error)

Cos performs a pointwise cos.

func Cube ¶ Uses

func Cube(a *Node) (*Node, error)

Cube performs a pointwise cube.

func DiagFlat ¶ Uses

func DiagFlat(a *Node) (*Node, error)

DiagFlat takes the flattened value and creates a diagonal matrix from it.

It is non-differentiable.

g := NewGraph()

// 2 dimensional
aV := tensor.New(tensor.WithShape(2, 2), tensor.WithBacking([]float64{1, 2, 3, 4}))
a := NodeFromAny(g, aV)
b, err := DiagFlat(a)
if err != nil {
    fmt.Println(err)
    return
}
m := NewTapeMachine(g)
if err := m.RunAll(); err != nil {
    fmt.Println(err)
    return
}
fmt.Printf("a:\n%v\n", a.Value())
fmt.Printf("b:\n%v\n", b.Value())

// 3 dimensional
aV = tensor.New(tensor.WithShape(2, 3, 2), tensor.WithBacking([]float64{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12}))
a = NodeFromAny(g, aV, WithName("a'"))
b2, err := DiagFlat(a)
if err != nil {
    fmt.Println(err)
    return
}
m = NewTapeMachine(g)
if err := m.RunAll(); err != nil {
    fmt.Println(err)
}

fmt.Printf("a:\n%v", a.Value())
fmt.Printf("b:\n%v\n", b2.Value())

// 1 dimensional
aV = tensor.New(tensor.WithShape(2), tensor.WithBacking([]float64{1, 2}))
a = NodeFromAny(g, aV, WithName("a''"))
b3, err := DiagFlat(a)
if err != nil {
    fmt.Println(err)
    return
}
m = NewTapeMachine(g)
if err := m.RunAll(); err != nil {
    fmt.Println(err)
}

fmt.Printf("a:\n%v\n", a.Value())
fmt.Printf("b:\n%v\n", b3.Value())

// Scalars

a = NodeFromAny(g, 100.0, WithName("aScalar"))
_, err = DiagFlat(a)
fmt.Println(err)

a:
⎡1  2⎤
⎣3  4⎦

b:
⎡1  0  0  0⎤
⎢0  2  0  0⎥
⎢0  0  3  0⎥
⎣0  0  0  4⎦

a:
⎡ 1   2⎤
⎢ 3   4⎥
⎣ 5   6⎦

⎡ 7   8⎤
⎢ 9  10⎥
⎣11  12⎦

b:
⎡ 1   0   0   0  ...  0   0   0   0⎤
⎢ 0   2   0   0  ...  0   0   0   0⎥
⎢ 0   0   3   0  ...  0   0   0   0⎥
⎢ 0   0   0   4  ...  0   0   0   0⎥
.
.
.
⎢ 0   0   0   0  ...  9   0   0   0⎥
⎢ 0   0   0   0  ...  0  10   0   0⎥
⎢ 0   0   0   0  ...  0   0  11   0⎥
⎣ 0   0   0   0  ...  0   0   0  12⎦

a:
[1  2]
b:
⎡1  0⎤
⎣0  2⎦

Cannot perform DiagFlat on a scalar equivalent node

func Div ¶ Uses

func Div(a, b *Node) (retVal *Node, err error)

Div is a shortcut function for HadamardDiv for scalar values. For matrix/tensor values, the matrix division operation is not yet handled, and will panic.

func Dropout ¶ Uses

func Dropout(x *Node, dropProb float64) (retVal *Node, err error)

Dropout is a convenience function to implement dropout. It uses randomly zeroes out a *Tensor with a probability drawn from a uniform distribution

func Eq ¶ Uses

func Eq(a, b *Node, retSame bool) (*Node, error)

Eq performs a pointwise eq operation. retSame indicates if the data type of the return value should be the same as the input data type. It defaults to Bool otherwise.

func Exp ¶ Uses

func Exp(a *Node) (*Node, error)

Exp performs a pointwise exp.

func Expm1 ¶ Uses

func Expm1(a *Node) (*Node, error)

Expm1 performs a pointwise expm1.

func Floor ¶ Uses

func Floor(a *Node) (*Node, error)

Floor performs a pointwise floor.

func GaussianRandomNode ¶ Uses

func GaussianRandomNode(g *ExprGraph, dt tensor.Dtype, mean, stdev float64, shape ...int) *Node

GaussianRandomNode creates an input node that has a random op so everytime the node is passed, random values will be plucked from a gaussian distribution with the mean and stdev provided. The type of the node depends on the shape passed in. To get a scalar value at run time, don't pass in any shapes

func GlobalAveragePool2D ¶ Uses

func GlobalAveragePool2D(x *Node) (*Node, error)

GlobalAveragePool2D consumes an input tensor X and applies average pooling across the values in the same channel. The expected input shape is BCHW where B is the batch size, C is the number of channels, and H and W are the height and the width of the data.

func Gt ¶ Uses

func Gt(a, b *Node, retSame bool) (*Node, error)

Gt performs a pointwise gt operation. retSame indicates if the data type of the return value should be the same as the input data type. It defaults to Bool otherwise.

func Gte ¶ Uses

func Gte(a, b *Node, retSame bool) (*Node, error)

Gte performs a pointwise gte operation. retSame indicates if the data type of the return value should be the same as the input data type. It defaults to Bool otherwise.

func HadamardDiv ¶ Uses

func HadamardDiv(a, b *Node) (*Node, error)

HadamardDiv performs a pointwise hadamarddiv operation.

func HadamardProd ¶ Uses

func HadamardProd(a, b *Node) (*Node, error)

HadamardProd performs a pointwise hadamardprod operation.

func Im2Col ¶ Uses

func Im2Col(n *Node, kernel, pad, stride, dilation tensor.Shape) (retVal *Node, err error)

Im2Col converts a BCHW image block to columns. The kernel, pad and stride parameter must be shape of size 2, no more no less This poor naming scheme clearly comes from matlab

func Inverse ¶ Uses

func Inverse(a *Node) (*Node, error)

Inverse performs a pointwise inverse.

func InverseSqrt ¶ Uses

func InverseSqrt(a *Node) (*Node, error)

InverseSqrt performs a pointwise inversesqrt.

func KeepDims ¶ Uses

func KeepDims(a *Node, expandLeft bool, fn func(a *Node) (*Node, error)) (*Node, error)

KeepDims is a function that ensures that input and output dimensions are the same though the shape may change.

The expandLeft flag in the function indicates if any shape expansion should be done leftwards or rightwards. For example, if fn() returns a tensor with a shape (3) and the desired dimension is 2, then if `expandLeft` is true the result will be `(1, 3)`. Otherwise the result will be `(3, 1)`.

At the moment, results that turn into scalars cannot have their dimensions kept - the semantics isn't well established yet and is a work in progress.

func LeakyRelu ¶ Uses

func LeakyRelu(x *Node, alpha float64) (*Node, error)

LeakyRelu returns a node whose underlying value is:

f(x) = alpha * x if x < 0
f(x) = x for x ⩾ 0

applied elementwise.

func Log ¶ Uses

func Log(a *Node) (*Node, error)

Log performs a pointwise log.

func Log1p ¶ Uses

func Log1p(a *Node) (*Node, error)

Log1p performs a pointwise log1p.

func Log2 ¶ Uses

func Log2(a *Node) (*Node, error)

Log2 performs a pointwise log2.

func LogSumExp ¶ Uses

func LogSumExp(a *Node, axis int) (retVal *Node, err error)

LogSumExp performs addition in the log domain

func Lt ¶ Uses

func Lt(a, b *Node, retSame bool) (*Node, error)

Lt performs a pointwise lt operation. retSame indicates if the data type of the return value should be the same as the input data type. It defaults to Bool otherwise.

func Lte ¶ Uses

func Lte(a, b *Node, retSame bool) (*Node, error)

Lte performs a pointwise lte operation. retSame indicates if the data type of the return value should be the same as the input data type. It defaults to Bool otherwise.

func Max ¶ Uses

func Max(a *Node, along ...int) (retVal *Node, err error)

Max performs a max() on the input and the provided axes.

func MaxPool1D ¶ Uses

func MaxPool1D(x *Node, kernel, pad, stride int) (*Node, error)

MaxPool1D applies a maxpool on the node x.

func MaxPool2D ¶ Uses

func MaxPool2D(x *Node, kernel tensor.Shape, pad, stride []int) (*Node, error)

MaxPool2D applies the kernel filter to the input node. The pad slice can have two different lengths.

- if len(pad) == 2, padding is assume to be symetric, and a padding is adding up *and* down to each dimension

paddedOutputH = pad[0] + inputH + pad[0]
paddedOutputW = pad[1] + inputW + pad[1]

- if len(pad) == 4, padding is explicit and can be asymmetric.

paddedOutputH = pad[0] + inputH + pad[1]
paddedOutputW = pad[2] + inputW + pad[3]

func Mean ¶ Uses

func Mean(a *Node, along ...int) (retVal *Node, err error)

Mean performs a mean() on the input and the provided axes.

func Mish ¶ Uses

func Mish(a *Node) (retVal *Node, err error)

Mish is a novel activation function that is self regularizing.

https://arxiv.org/abs/1908.08681

func Mul ¶ Uses

func Mul(a, b *Node) (retVal *Node, err error)

Mul is the general handler for multiplication of nodes. It is extremely overloaded. Only use if you know what you're doing

If any of the nodes are ScalarType, then it'll be redirected to HadamardProd() instead If the nodes are both vectors (that is, have a shape of (x, 1) or (1, x)), then the operator used will be a vectorDot If only one of the nodes is a vector, then the operator used will be a matrix-vector multiplication will be used, and most importantly, a transpose will be used (when necessary) If both nodes are matrices, then well, matrix multiplication will be done

func Must ¶ Uses

func Must(n *Node, err error, opts ...NodeConsOpt) *Node

Must indicates a node must be created. If there isn't a node created, or there was an error, it subsumes the error, and immediately panics

func Ne ¶ Uses

func Ne(a, b *Node, retSame bool) (*Node, error)

Ne performs a pointwise ne operation. retSame indicates if the data type of the return value should be the same as the input data type. It defaults to Bool otherwise.

func Neg ¶ Uses

func Neg(a *Node) (*Node, error)

Neg performs a pointwise neg.

func NewConstant ¶ Uses

func NewConstant(v interface{}, opts ...NodeConsOpt) *Node

NewConstant takes in any reasonable value and makes it a constant node.

func NewMatrix ¶ Uses

func NewMatrix(g *ExprGraph, t tensor.Dtype, opts ...NodeConsOpt) *Node

NewMatrix creates a Node representing a variable that holds a matrix (nxm)

func NewScalar ¶ Uses

func NewScalar(g *ExprGraph, t tensor.Dtype, opts ...NodeConsOpt) *Node

NewScalar creates a Node representing a variable that holds a scalar value

func NewTensor ¶ Uses

func NewTensor(g *ExprGraph, t tensor.Dtype, dims int, opts ...NodeConsOpt) *Node

NewTensor creates a Node representing a variable that holds a tensor (any n-dimensional array with dimensions greater than 2)

func NewUniqueNode ¶ Uses

func NewUniqueNode(opts ...NodeConsOpt) *Node

NewUniqueNode creates a new unique node in a graph. If no graph was specified in the construction options then it will just return a graphless node.

func NewVector ¶ Uses

func NewVector(g *ExprGraph, t tensor.Dtype, opts ...NodeConsOpt) *Node

NewVector creates a Node representing a variable that holds a vector (nx1 matrix)

func NodeFromAny ¶ Uses

func NodeFromAny(g *ExprGraph, any interface{}, opts ...NodeConsOpt) *Node

NodeFromAny creates a Node from a tensor.Tensor, automatically filling in shape and type info

func Norm ¶ Uses

func Norm(a *Node, axis, p int) (retVal *Node, err error)

Norm returns the p-norm of a Value. Use p=2 if you want to use unordered norms.

This is a simpler version of the norms found in the Tensor package, which specializes and optimizes even more (well, given it's adapted from Numpy, it is clearly way more optimized)

func OneHotVector ¶ Uses

func OneHotVector(id, classes int, t tensor.Dtype, opts ...NodeConsOpt) *Node

OneHotVector creates a node representing a one hot vector

func OuterProd ¶ Uses

func OuterProd(a, b *Node) (retVal *Node, err error)

OuterProd returns a Node representing the outer product of two vectors. This function will return an error if both input nodes are not vectors

func Pow ¶ Uses

func Pow(a, b *Node) (*Node, error)

Pow performs a pointwise pow operation.

func Ravel ¶ Uses

func Ravel(n *Node) (retVal *Node, err error)

Ravel flattens the given node and returns the new node

func Read ¶ Uses

func Read(n *Node, into *Value) (retVal *Node)

Read allows for extraction of the value of the *Node at runtime into a Value. To achieve this, a pointer to a Value (*Value) is passed into this function, not a Value. The 'into' value remains nil until the execution of the graph (via a call to the Run() methods of the VM)

func Rectify ¶ Uses

func Rectify(x *Node) (retVal *Node, err error)

Rectify is a convenience function for creating rectified linear units activation functions. This function uses ⩾, which is the canonical version. If you want to use >, you can create your own by just following this.

func ReduceAdd ¶ Uses

func ReduceAdd(nodes Nodes, opts ...NodeConsOpt) (retVal *Node, err error)

ReduceAdd takes a slice of *Nodes, and folds them into one by adding

func ReduceMul ¶ Uses

func ReduceMul(nodes Nodes, opts ...NodeConsOpt) (retVal *Node, err error)

ReduceMul is like foldl(*, nodes)

func Reshape ¶ Uses

func Reshape(n *Node, to tensor.Shape) (retVal *Node, err error)

Reshape reshapes a node and returns a new node with the new shape

func Set ¶ Uses

func Set(a, b *Node) (retVal *Node)

Set is the equivalent of doing this:

a = b

where a and b are both variables

func Sigmoid ¶ Uses

func Sigmoid(a *Node) (*Node, error)

Sigmoid performs a pointwise sigmoid.

func Sign ¶ Uses

func Sign(a *Node) (*Node, error)

Sign performs a pointwise sign.

func Sin ¶ Uses

func Sin(a *Node) (*Node, error)

Sin performs a pointwise sin.

func SizeOf ¶ Uses

func SizeOf(axis int, x *Node) (retVal *Node, err error)

SizeOf returns the size of a value along an axis

func Slice ¶ Uses

func Slice(n *Node, slices ...tensor.Slice) (retVal *Node, err error)

Slice slices a *Node. For T[:] slices, pass in nil. Will error out if node's type is not a Tensor

func SoftMax ¶ Uses

func SoftMax(x *Node, axis ...int) (*Node, error)

SoftMax implements the softmax operation. The softmax operation is a stable operation.

g := NewGraph()
t := tensor.New(tensor.WithShape(2, 3), tensor.WithBacking([]float64{1, 3, 2, 3, 2, 1}))
u := t.Clone().(*tensor.Dense)
v := tensor.New(tensor.WithShape(2, 2, 3), tensor.WithBacking([]float64{
    1, 3, 2,
    4, 2, 1,

    3, 5, 3,
    2, 1, 5,
}))

a := NodeFromAny(g, t, WithName("a"))
b := NodeFromAny(g, u, WithName("b"))
c := NodeFromAny(g, v, WithName("c"))

sm1 := Must(SoftMax(a))
sm0 := Must(SoftMax(b, 0))
sm := Must(SoftMax(c))
m := NewTapeMachine(g)
if err := m.RunAll(); err != nil {
    panic(err)
}

fmt.Printf("a:\n%v\nsoftmax(a) - along last axis (default behaviour):\n%1.2f", a.Value(), sm1.Value())
fmt.Printf("b:\n%v\nsoftmax(b) - along axis 0:\n%1.2f", b.Value(), sm0.Value())

tmp := fmt.Sprintf("c %v:\n%v\nsoftmax(c) - along last axis (default behaviour) %v:\n%1.2f", c.Value().Shape(), c.Value(), sm.Value().Shape(), sm.Value())

fmt.Println(strings.Replace(tmp, "\n\n\n", "\n\n", -1))

// the requirement to use tmp and strings.Replace is because when Go runs example tests, it strips excess newlines.

a:
⎡1  3  2⎤
⎣3  2  1⎦

softmax(a) - along last axis (default behaviour):
⎡0.09  0.67  0.24⎤
⎣0.67  0.24  0.09⎦
b:
⎡1  3  2⎤
⎣3  2  1⎦

softmax(b) - along axis 0:
⎡0.12  0.73  0.73⎤
⎣0.88  0.27  0.27⎦
c (2, 2, 3):
⎡1  3  2⎤
⎣4  2  1⎦

⎡3  5  3⎤
⎣2  1  5⎦

softmax(c) - along last axis (default behaviour) (2, 2, 3):
⎡0.09  0.67  0.24⎤
⎣0.84  0.11  0.04⎦

⎡0.11  0.79  0.11⎤
⎣0.05  0.02  0.94⎦

func Softplus ¶ Uses

func Softplus(a *Node) (*Node, error)

Softplus performs a pointwise softplus.

func Sparsemax ¶ Uses

func Sparsemax(x *Node, axes ...int) (*Node, error)

Sparsemax - implements the sparsemax operation described here: http://proceedings.mlr.press/v48/martins16.pdf

func Sqrt ¶ Uses

func Sqrt(a *Node) (*Node, error)

Sqrt performs a pointwise sqrt.

func Square ¶ Uses

func Square(a *Node) (*Node, error)

Square performs a pointwise square.

func Sub ¶ Uses

func Sub(a, b *Node) (*Node, error)

Sub performs a pointwise sub operation.

func Sum ¶ Uses

func Sum(a *Node, along ...int) (retVal *Node, err error)

Sum performs a sum() on the input and the provided axes.

func Tanh ¶ Uses

func Tanh(a *Node) (*Node, error)

Tanh performs a pointwise tanh.

func Tensordot ¶ Uses

func Tensordot(aAxes []int, bAxes []int, a, b *Node) (retVal *Node, err error)

Tensordot performs a tensor contraction of a and b along specified axes.

// Scalars
g := NewGraph()
a := NewScalar(g, Float64, WithValue(2.0), WithName("a"))
b := NewScalar(g, Float64, WithValue(21.0), WithName("b"))
c, err := Tensordot([]int{0}, []int{0}, a, b)
if err != nil {
	fmt.Printf("Cannot call Tensordot. Error: %v\n", err)
	return
}

vm := NewTapeMachine(g)
if err := vm.RunAll(); err != nil {
	fmt.Printf("Cannot perform scalars. Error %v\n", err)
}
fmt.Printf("c: %v (%v) of %v", c.Value(), c.Value().Dtype(), c.Value().Shape())

// Output:
//...

g := NewGraph()
a := NewVector(g, Float64, WithName("a"), WithShape(2), WithInit(RangedFrom(2)))
b := NewVector(g, Float64, WithName("b"), WithShape(2), WithInit(RangedFrom(21)))

c, err := Tensordot([]int{0}, []int{0}, a, b)
if err != nil {
    fmt.Printf("Cannot call Tensordot. Error: %v\n", err)
    return
}

vm := NewTapeMachine(g)
if err := vm.RunAll(); err != nil {
    fmt.Printf("Cannot perform tensordot on vectors. Error %v\n", err)
}
fmt.Printf("a %v b %v ", a.Value(), b.Value())
fmt.Printf("c: %v (%v) of %v", c.Value(), c.Type(), c.Value().Shape())

a [2  3] b [21  22] c: [108] (float64) of (1)

func Transpose ¶ Uses

func Transpose(n *Node, axes ...int) (retVal *Node, err error)

Transpose performs a transpose on the input and provided permutation axes.

func UniformRandomNode ¶ Uses

func UniformRandomNode(g *ExprGraph, dt tensor.Dtype, low, high float64, shape ...int) *Node

UniformRandomNode creates an input node that has a random op so everytime the node is passed, random values will be plucked from a uniform distribution. The type of the node depends on the shape passed in. To get a scalar value at run time, don't pass in any shapes

func Upsample2D ¶ Uses

func Upsample2D(x *Node, scale int) (*Node, error)

Upsample2D - simply upscaling Tensor by scale factor.

1, 2
3, 4
converts to
1,1,2,2
1,1,2,2
3,3,4,4,
3,3,4,4,

func YOLOv3 ¶ Uses

func YOLOv3(input *Node, anchors []float32, masks []int, netSize, numClasses int, ignoreTresh float32, targets ...*Node) (*Node, error)

YOLOv3 https://arxiv.org/abs/1804.02767

func (*Node) Clone ¶ Uses

func (n *Node) Clone() (retVal interface{})

Clone clones the node. There are some caveats:

- the graph is not copied over - the node essentially does not belong to a collection
- there is no ID
- the children are not cloned

func (*Node) CloneTo ¶ Uses

func (n *Node) CloneTo(g *ExprGraph) *Node

CloneTo clones the node into a new graph. If CloneTo() is called on the same graph as the n, it will return n. The reason this is done is because at any given time, every node should be unique in the *ExprGraph.

TODO: clone children as well (this means that CloneTo() is only currently suitable fo input nodes)

func (*Node) DataSize ¶ Uses

func (n *Node) DataSize() int

func (*Node) Deriv ¶ Uses

func (n *Node) Deriv() *Node

func (*Node) DerivOf ¶ Uses

func (n *Node) DerivOf() Nodes

func (*Node) Device ¶ Uses

func (n *Node) Device() Device

Device returns the device the data will be on

func (*Node) Dims ¶ Uses

func (n *Node) Dims() int

Dims indicates how many dimensions the node's result has

func (*Node) Dtype ¶ Uses

func (n *Node) Dtype() tensor.Dtype

Dtype returns the dtype of the node

func (*Node) Err ¶ Uses

func (n *Node) Err() error

Err always returns nil. However, this method is implemented to enable nicer composition of functions

func (*Node) Grad ¶ Uses

func (n *Node) Grad() (Value, error)

Grad returns the gradient if there is one.

func (*Node) GradOnDevice ¶ Uses

func (n *Node) GradOnDevice(dev Device, extern External) (retVal Value, allocOnExtern bool, err error)

GradOnDevice gets the gradient value of the node as a Value but on the desired device. In this build the device is always CPU, so it's equivalent to calling .Grad()

func (*Node) Graph ¶ Uses

func (n *Node) Graph() *ExprGraph

Graph returns the graph of the node

func (*Node) Groups ¶ Uses

func (n *Node) Groups() encoding.Groups

Groups to fulfil the encoding Grouper interface

func (*Node) Hashcode ¶ Uses

func (n *Node) Hashcode() uint32

Hashcode provides the hash for the tree, assuming that the node is the root of the tree. Original implementation was here by Vatine (who's apparently 80 years old and using SO!?!):

http://stackoverflow.com/questions/1988665/hashing-a-tree-structure

func (*Node) ID ¶ Uses

func (n *Node) ID() int64

ID returns the ID of the node. This satisfies the gonum/graph.Node interface

func (*Node) IsColVec ¶ Uses

func (n *Node) IsColVec() bool

IsColVec indicates if a node represents a Column Vector. This is based on the type of the node, not the actual value associated with the node

func (*Node) IsMatrix ¶ Uses

func (n *Node) IsMatrix() bool

IsMatrix indicates if a node represents a matrix. This is based on the type of the node, not the actual value associated with the node

func (*Node) IsRowVec ¶ Uses

func (n *Node) IsRowVec() bool

IsRowVec indicates if a node represents a Row Vector. This is based on the type of the node, not the actual value associated with the node

func (*Node) IsScalar ¶ Uses

func (n *Node) IsScalar() bool

IsScalar indicates if a node represents a a scalar value. This is based on the type of the node, not the actual value associated with the node

func (*Node) IsVar ¶ Uses

func (n *Node) IsVar() bool

IsVar returns true if the node represents a differentiable variable (i.e. it's an argument to the function that is not a statement)

func (*Node) IsVec ¶ Uses

func (n *Node) IsVec() bool

IsVec returns whether this node is a vector

func (*Node) IsVector ¶ Uses

func (n *Node) IsVector() bool

IsVector indicates if a node represents a vector value. This is based on the type of the node, not the actual value associated with the node

func (*Node) Name ¶ Uses

func (n *Node) Name() string

Name returns the name of the node. If a name was specified and it is too long, the short name will be used instead (except in inputs)

The short name is typically of the form: OpName(%1, %2 ...), making it read more like a function call

func (*Node) Node ¶ Uses

func (n *Node) Node() *Node

Node returns itself. This sorts of monoidal patterns are useful for compositions via interfaces.

func (*Node) Nodes ¶ Uses

func (n *Node) Nodes() Nodes

Nodes returns n as a slice of *Node. Again, this is mostly useful for interfaces

func (*Node) Op ¶ Uses

func (n *Node) Op() Op

Op returns the Op of the node

func (*Node) RestrictedToDot ¶ Uses

func (n *Node) RestrictedToDot(up, down int) string

RestrictedToDot prints the graphviz compatible string but does not print the entire tree up and down indicates how many levels to look up, and how many levels to look down

func (*Node) Shape ¶ Uses

func (n *Node) Shape() tensor.Shape

Shape returns the shape of the node

func (*Node) Strides ¶ Uses

func (n *Node) Strides() []int

Strides returns the strides of the value of the node

func (*Node) String ¶ Uses

func (n *Node) String() string

String() implements the fmt.Stringer interface

func (*Node) ToDot ¶ Uses

func (n *Node) ToDot() string

ToDot returns the graph as a graphviz compatible string. DEPRECATED: This function will be removed in the next release, please use the encoding/dot package

func (*Node) Type ¶ Uses

func (n *Node) Type() hm.Type

Type returns the type of the node

func (*Node) Value ¶ Uses

func (n *Node) Value() Value

Value returns the valuse bound to the node. May return nil

func (*Node) ValueOnDevice ¶ Uses

func (n *Node) ValueOnDevice(dev Device, extern External) (retVal Value, allocOnExtern bool, err error)

ValueOnDevice gets the value of the node as a Value but on the desired device. In this build the device is always CPU, so it's equivalent to calling .Value()

func (*Node) WriteHash ¶ Uses

func (n *Node) WriteHash(h hash.Hash32)

WriteHash writes the hash to the provided Hash32.

type NodeConsOpt ¶ Uses

type NodeConsOpt func(*Node)

NodeConsOpt is a function that provides construction options for any Node.

func In ¶ Uses

func In(g *ExprGraph) NodeConsOpt

In is a node construction option to set a node's graph. A `*Node`'s graph is immutable. If the graph has already been set, a check will be made that the specifiec *Graph and the *Graph set in *Node are the same. If they are not, the function will panic/

func WithChildren ¶ Uses

func WithChildren(children Nodes) NodeConsOpt

WithChildren sets the children of a node to the specified chidren. This construction option does NOT check if existing children exists, and will overwrite the existing children.

func WithGrad ¶ Uses

func WithGrad(any interface{}) NodeConsOpt

WithGrad is a node construction option that binds the value to the *Node. This function may panic if:

- There isn't already a value associated with the node (.boundTo == nil)
- The type of the Value does not match the value of the node.

func WithGroupName ¶ Uses

func WithGroupName(name string) NodeConsOpt

WithGroupName is a node construction option to group a *Node within a particular group. This option is useful for debugging with graphs. This function is deprecated and will proabably be remove in the next version.

func WithInit ¶ Uses

func WithInit(fn InitWFn) NodeConsOpt

WithInit is a node construction option to initialize a *Node with the InitWFn provided.

func WithName ¶ Uses

func WithName(name string) NodeConsOpt

WithName is a node construction option that gives the *Node the provided name. This is especially useful in debugging graphs.

func WithOp ¶ Uses

func WithOp(op Op) NodeConsOpt

WithOp is a node construction option to set a node's Op to the specified Op. `Op`s in `*Node`s are immutable once set and cannot be changed. If the node already has an Op specified a check will be made to see if the provided Op and the one already specified in the `*Node` is the same - do note that comparison of Ops is done using the `Hashcode()` method of Ops, and hash collisions MAY occur - If both ops are different, this function will panic.

func WithShape ¶ Uses

func WithShape(shp ...int) NodeConsOpt

WithShape is a node construction option to initialize a *Node with a particular shape. This function panics if the shape's dimensions do not match the specified dimensions of the *Node.

func WithType ¶ Uses

func WithType(t hm.Type) NodeConsOpt

WithType is a node construction option to set a node to the specified type. Types in *Node are immutable once set. If the type has already been specified in the node, a check will be made to see if the both types are the same. If it isn't, it will panic.

func WithValue ¶ Uses

func WithValue(any interface{}) NodeConsOpt

WithValue is a node construction option that binds the value to the *Node. This function may panic if:

- Gorgonia was unable to convert interface{} into a Value.
- The type of the Value does not match the type of the nodes.

type NodeSet ¶ Uses

type NodeSet map[*Node]struct{}

NodeSet is the primary type that represents a set

func NewNodeSet ¶ Uses

func NewNodeSet(a ...*Node) NodeSet

NewNodeSet creates and returns a reference to an empty set.

func (NodeSet) Add ¶ Uses

func (set NodeSet) Add(i *Node) bool

Add adds an item to the current set if it doesn't already exist in the set.

func (NodeSet) Cardinality ¶ Uses

func (set NodeSet) Cardinality() int

Cardinality returns how many items are currently in the set.

func (*NodeSet) Clear ¶ Uses

func (set *NodeSet) Clear()

Clear clears the entire set to be the empty set.

func (NodeSet) Clone ¶ Uses

func (set NodeSet) Clone() NodeSet

Clone returns a clone of the set. Does NOT clone the underlying elements.

func (NodeSet) Contains ¶ Uses

func (set NodeSet) Contains(i *Node) bool

Contains determines if a given item is already in the set.

func (NodeSet) ContainsAll ¶ Uses

func (set NodeSet) ContainsAll(i ...*Node) bool

ContainsAll determines if the given items are all in the set

func (NodeSet) Difference ¶ Uses

func (set NodeSet) Difference(other NodeSet) NodeSet

Difference returns a new set with items in the current set but not in the other set

func (NodeSet) Equal ¶ Uses

func (set NodeSet) Equal(other NodeSet) bool

Equal determines if two sets are equal to each other. If they both are the same size and have the same items they are considered equal. Order of items is not relevant for sets to be equal.

func (NodeSet) Intersect ¶ Uses

func (set NodeSet) Intersect(other NodeSet) NodeSet

Intersect returns a new set with items that exist only in both sets.

func (NodeSet) IsSubset ¶ Uses

func (set NodeSet) IsSubset(other NodeSet) bool

IsSubset determines if every item in the other set is in this set.

func (NodeSet) IsSuperset ¶ Uses

func (set NodeSet) IsSuperset(other NodeSet) bool

IsSuperset determines if every item of this set is in the other set.

func (NodeSet) Iter ¶ Uses

func (set NodeSet) Iter() <-chan *Node

Iter returns a channel of type *Node that you can range over.

func (NodeSet) Remove ¶ Uses

func (set NodeSet) Remove(i *Node)

Remove allows the removal of a single item in the set.

func (NodeSet) SymmetricDifference ¶ Uses

func (set NodeSet) SymmetricDifference(other NodeSet) NodeSet

SymmetricDifference returns a new set with items in the current set or the other set but not in both.

func (NodeSet) ToSlice ¶ Uses

func (set NodeSet) ToSlice() Nodes

ToSlice returns the elements of the current set as a slice

func (NodeSet) Union ¶ Uses

func (set NodeSet) Union(other NodeSet) NodeSet

Union returns a new set with all items in both sets.

type Nodes ¶ Uses

type Nodes []*Node

Nodes is a slice of nodes, but it also acts as a set of nodes by implementing the Sort interface

func Backpropagate ¶ Uses

func Backpropagate(outputs, gradOutputs, wrt Nodes) (retVal Nodes, err error)

Backpropagate backpropagates errors by performing reverse-mode symbolic differentiation, starting from the outputs, and working its way towads the inputs.

This is the rough algorithm:

1. Filter out nodes that are unreachable
2. Forwards analysis, where a list of nodes affecting the output is added to consideration
3. Backwards analysis, where a list of nodes affected by differentiating the output are added to the consideration
4. If there is a difference in both sets, it will cause an error (both sets should be the same)
5. Traverse the graph from output towards input. On each visit, perform the symbolic differentiation

For most cases, Grad() should be used instead of Backpropagate(), as Grad() performs several checks which would be the general use case, before calling Backpropagate()

func Grad ¶ Uses

func Grad(cost *Node, WRTs ...*Node) (retVal Nodes, err error)

Grad takes a scalar cost node and a list of with-regards-to, and returns the gradient

func NodesFromInputs ¶ Uses

func NodesFromInputs(xs ...Input) (Nodes, error)

NodesFromInputs creates a Nodes from a list of Input.

func Sort ¶ Uses

func Sort(g *ExprGraph) (sorted Nodes, err error)

Sort topologically sorts a ExprGraph: root of graph will be first nodes are sorted using gonum's SortStabilized function.

see https://godoc.org/gonum.org/v1/gonum/graph/topo#SortStabilized for more info

func Unconcat ¶ Uses

func Unconcat(a *Node, along int, n int) (Nodes, error)

Unconcat is the opposite of the built in concat function TODO: port this back to Gorgonia and use Gorgonia's sli instead

g := NewGraph()
x := NewTensor(g, Float64, 4, WithShape(2, 3, 4, 5), WithInit(RangedFrom(0)), WithName("x"))
y := NewTensor(g, Float64, 4, WithShape(2, 3, 4, 5), WithInit(RangedFrom(120)), WithName("y"))

z, err := Concat(2, x, y)
if err != nil {
    panic(err)
}

unconcats, err := Unconcat(z, 2, 2)
if err != nil {
    panic(err)
}
a, b := unconcats[0], unconcats[1]

m := NewTapeMachine(g)
if err := m.RunAll(); err != nil {
    panic(err)
}
tmp := fmt.Sprintf("a %v\n%v\nb %v\n%v", a.Value().Shape(), a.Value(), b.Value().Shape(), b.Value())
fmt.Println(strings.Replace(tmp, "\n\n", "\n", -1))

a (2, 3, 4, 5)
⎡  0    1    2    3    4⎤
⎢  5    6    7    8    9⎥
⎢ 10   11   12   13   14⎥
⎣ 15   16   17   18   19⎦

⎡ 20   21   22   23   24⎤
⎢ 25   26   27   28   29⎥
⎢ 30   31   32   33   34⎥
⎣ 35   36   37   38   39⎦

⎡ 40   41   42   43   44⎤
⎢ 45   46   47   48   49⎥
⎢ 50   51   52   53   54⎥
⎣ 55   56   57   58   59⎦

⎡ 60   61   62   63   64⎤
⎢ 65   66   67   68   69⎥
⎢ 70   71   72   73   74⎥
⎣ 75   76   77   78   79⎦

⎡ 80   81   82   83   84⎤
⎢ 85   86   87   88   89⎥
⎢ 90   91   92   93   94⎥
⎣ 95   96   97   98   99⎦

⎡100  101  102  103  104⎤
⎢105  106  107  108  109⎥
⎢110  111  112  113  114⎥
⎣115  116  117  118  119⎦

b (2, 3, 4, 5)
⎡120  121  122  123  124⎤
⎢125  126  127  128  129⎥
⎢130  131  132  133  134⎥
⎣135  136  137  138  139⎦

⎡140  141  142  143  144⎤
⎢145  146  147  148  149⎥
⎢150  151  152  153  154⎥
⎣155  156  157  158  159⎦

⎡160  161  162  163  164⎤
⎢165  166  167  168  169⎥
⎢170  171  172  173  174⎥
⎣175  176  177  178  179⎦

⎡180  181  182  183  184⎤
⎢185  186  187  188  189⎥
⎢190  191  192  193  194⎥
⎣195  196  197  198  199⎦

⎡200  201  202  203  204⎤
⎢205  206  207  208  209⎥
⎢210  211  212  213  214⎥
⎣215  216  217  218  219⎦

⎡220  221  222  223  224⎤
⎢225  226  227  228  229⎥
⎢230  231  232  233  234⎥
⎣235  236  237  238  239⎦

func UnstableSort ¶ Uses

func UnstableSort(g *ExprGraph) (sorted Nodes, err error)

UnstableSort performs a topological sort of the directed graph g returning the 'from' to 'to' sort order. If a topological ordering is not possible, an Unorderable error is returned listing cyclic components in g with each cyclic component's members sorted by ID. When an Unorderable error is returned, each cyclic component's topological position within the sorted nodes is marked with a nil graph.Node.

func (Nodes) Add ¶ Uses

func (ns Nodes) Add(n *Node) Nodes

Add adds to set

func (Nodes) AllSameGraph ¶ Uses

func (ns Nodes) AllSameGraph() bool

AllSameGraph returns true if all the nodes in the slice belong to the same graph. Note that constants do not have to belong to the same graph.

func (Nodes) Contains ¶ Uses

func (ns Nodes) Contains(want *Node) bool

Contains checks if the wanted node is in the set

func (Nodes) Difference ¶ Uses

func (ns Nodes) Difference(other Nodes) Nodes

Difference is ns - other. Bear in mind it is NOT commutative

func (Nodes) Equals ¶ Uses

func (ns Nodes) Equals(other Nodes) bool

Equals returns true if two Nodes are the same

func (Nodes) Err ¶ Uses

func (ns Nodes) Err() error

Err returns nil always

func (Nodes) Format ¶ Uses

func (ns Nodes) Format(s fmt.State, c rune)

Format implements fmt.Formatter, which allows Nodes to be differently formatted depending on the verbs

func (Nodes) Intersect ¶ Uses

func (ns Nodes) Intersect(other Nodes) Nodes

Intersect performs an intersection with other Nodes

func (Nodes) Len ¶ Uses

func (ns Nodes) Len() int

func (Nodes) Less ¶ Uses

func (ns Nodes) Less(i, j int) bool

func (Nodes) Node ¶ Uses

func (ns Nodes) Node() *Node

Node returns nil. Always. This is bound to cause a panic somewhere if an program is not using it correctly. The reason for implementing this is so that it may fulfil common interfaces.

func (Nodes) Nodes ¶ Uses

func (ns Nodes) Nodes() Nodes

Nodes returns itself. This is useful for interfaces

func (Nodes) Set ¶ Uses

func (ns Nodes) Set() Nodes

Set returns a uniquifies slice. It mutates the slice.

func (Nodes) Swap ¶ Uses

func (ns Nodes) Swap(i, j int)

type Op ¶ Uses

type Op interface {

    // Arity returns the number of inputs the Op expects. -1 indicates that it's n-ary and will be determined at runtime
    Arity() int

    // Informs the type of the Op (not the node). This will be used by the type system to infer the final type of the node
    Type() hm.Type

    // returns the output shape as a function of the inputs
    InferShape(...DimSizer) (tensor.Shape, error)

    // executes the op
    Do(...Value) (Value, error)

    // indicates if the Op will return a pointer (allowing possible inplace edits) or by value
    // if it's false, the return value of the Op will be a copy of its input
    ReturnsPtr() bool

    // Does this op potentially call external (cgo or cuda) functions (thereby requiring extra overhead for Go's trampolining thing)
    CallsExtern() bool

    // overwriteInput() is a method which states which input the output will be overwriting.
    // This allows for some efficiency gains as the underlying arrays wouldn't have to be re-allocated.
    // The method returns an int instead of a bool because potentially different operations may be allowed
    // to overwrite certain inputs. For example, consider an operation to increment a value:
    // the IncrementOp would be a unary operator, and assuming we would like to overwrite the input,
    // the retVal of overwriteInput() will be 0 (inputs[0]).
    // -1 is returned if overwriting of input is disallowed
    OverwritesInput() int

    /* Other methods */
    WriteHash(h hash.Hash)
    Hashcode() uint32
    fmt.Stringer
}

An Op is a symbolic representation of an operation Think of them as functions, taking an input (or multiple), and outputting something

All Ops have type signatures that look like this:

OpName :: (Floats a) ⇒ Tensor a → Tensor a → Tensor a

type RMSPropSolver ¶ Uses

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

RMSPropSolver is a solver that implements Geoffrey Hinton's RMSProp gradient descent optimization algorithm. http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf

func NewRMSPropSolver ¶ Uses

func NewRMSPropSolver(opts ...SolverOpt) *RMSPropSolver

NewRMSPropSolver creates an RMSProp solver with these default values:

eta (learn rate)	  : 0.001
eps (smoothing factor): 1e-8
rho (decay factor)    : 0.999

func (*RMSPropSolver) Step ¶ Uses

func (s *RMSPropSolver) Step(model []ValueGrad) (err error)

Step steps through each node in the model and applies the RMSProp gradient descent algorithm on the value.

This function will error out if the nodes do not have an associated Grad value.

type ReductionOp ¶ Uses

type ReductionOp interface {
    Op

    IsReduction() bool
}

ReductionOp changes the shape of the node

type Result ¶ Uses

type Result interface {
    Input
    Errer
}

Result is either a Node or Nodes or error. It's a poor man's sum types and it's not sealed for good reason

func LiftResult ¶ Uses

func LiftResult(a Input, err error) Result

LiftResult creates a Result from a Input and error pair. If the error is not nil, the Input is discarded.

The usual use case is in a function that returns a `(*Node, error)`. e.g LiftResult(Add(a, b))

type SDOp ¶ Uses

type SDOp interface {
    Op

    // DiffWRT indicates if the op is differentiable with regards to the given number of inputs
    // returns []bool to indicate which input it is differentiable to
    DiffWRT(inputs int) []bool

    // SymDiff symbolically differentiates the op
    SymDiff(inputs Nodes, output, grad *Node) (retVal Nodes, err error)
}

A SDOp is an Op that supports symbolic differentiation

type Scalar ¶ Uses

type Scalar interface {
    Value
    // contains filtered or unexported methods
}

Scalar represents a scalar(non-array-based) value. Do note that it's the pointers of the scalar types (F64, F32, etc) that implement the Scalar interface. The main reason is primarily due to optimizations with regards to memory allocation and copying for device interoperability.

type Solver ¶ Uses

type Solver interface {
    Step([]ValueGrad) error
}

Solver is anything that does gradient updates. The name solvers is stolen from Caffe. A much shorter name than GradientUpdaters

type SolverOpt ¶ Uses

type SolverOpt func(s Solver)

SolverOpt is a function that provides construction options for a Solver

func WithBatchSize ¶ Uses

func WithBatchSize(batch float64) SolverOpt

WithBatchSize sets the batch size for the solver. Currently only Adam and Vanilla (basic SGD) has batch size support

func WithBeta1 ¶ Uses

func WithBeta1(beta1 float64) SolverOpt

WithBeta1 sets the beta1 param of the solver. Only works with Adam

func WithBeta2 ¶ Uses

func WithBeta2(beta2 float64) SolverOpt

WithBeta2 sets the beta1 param of the solver. Only works with Adam

func WithClip ¶ Uses

func WithClip(clip float64) SolverOpt

WithClip clips the gradient if it gets too crazy. By default all solvers do not have any clips attached

func WithEps ¶ Uses

func WithEps(eps float64) SolverOpt

WithEps sets the smoothing factor for the solver.

func WithL1Reg ¶ Uses

func WithL1Reg(l1reg float64) SolverOpt

WithL1Reg adds a L1 regularization parameter to the solver. By default, the solvers do not use any regularization param

func WithL2Reg ¶ Uses

func WithL2Reg(l2reg float64) SolverOpt

WithL2Reg adds a L2 regularization parameter to the solver. By default, the solvers do not use any regularization param

func WithLearnRate ¶ Uses

func WithLearnRate(eta float64) SolverOpt

WithLearnRate sets the learn rate or step size for the solver.

func WithMomentum ¶ Uses

func WithMomentum(momentum float64) SolverOpt

WithMomentum sets the momentum of the solver. It is a no-op is the solver's type is not Momentum

func WithRho ¶ Uses

func WithRho(rho float64) SolverOpt

WithRho sets the decay parameter of the RMSProp solver

type StandardEngine ¶ Uses

type StandardEngine struct {
    tensor.StdEng
}

StandardEngine is the default CPU engine for gorgonia

func (StandardEngine) Transpose ¶ Uses

func (e StandardEngine) Transpose(a tensor.Tensor, expStrides []int) error

Transpose tensor a according to expStrides

type SymDiffError ¶ Uses

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

SymDiffError provides the context at which an error occurred

func (SymDiffError) Error ¶ Uses

func (err SymDiffError) Error() string

func (SymDiffError) Grad ¶ Uses

func (err SymDiffError) Grad() *Node

Grad returns a specific grad involved in the error

func (SymDiffError) Grads ¶ Uses

func (err SymDiffError) Grads() map[*Node]Nodes

Grads returns the grads involved in the error

func (SymDiffError) Node ¶ Uses

func (err SymDiffError) Node() *Node

Node returns a specific node involved in the error

func (SymDiffError) Nodes ¶ Uses

func (err SymDiffError) Nodes() Nodes

Nodes returns the nodes involved in the error

type Tensor ¶ Uses

type Tensor interface {
    // info about the ndarrayN
    Shape() tensor.Shape
    Strides() []int
    Dtype() tensor.Dtype
    Dims() int
    Size() int
    DataSize() int

    // type overloading methods
    IsScalar() bool
    ScalarValue() interface{}

    // engine/memory related stuff
    // all Tensors should be able to be expressed of as a slab of memory
    // Note: the size of each element can be acquired by T.Dtype().Size()
    Engine() tensor.Engine      // Engine can be nil
    MemSize() uintptr           // the size in memory
    Uintptr() uintptr           // the pointer to the first element, as a uintptr
    Pointer() unsafe.Pointer    // the pointer to the first elemment as a unsafe.Ponter
    IsNativelyAccessible() bool // Can Go access the memory
    IsManuallyManaged() bool    // Must Go manage the memory
}

Tensor is an interface that describes an ndarray

type TensorType ¶ Uses

type TensorType struct {
    Dims int // dims

    Of  hm.Type
}

TensorType is a type constructor for tensors.

Think of it as something like this:

data Tensor a = Tensor d a

The shape of the Tensor is not part of TensorType. Shape checking is relegated to the dynamic part of the program run

func (TensorType) Apply ¶ Uses

func (t TensorType) Apply(sub hm.Subs) hm.Substitutable

Apply applies the substitutions on the types. Satisfies the hm.Type interface.

func (TensorType) Eq ¶ Uses

func (t TensorType) Eq(other hm.Type) bool

Eq is the equality function of this type. The type of Tensor has to be the same, and for now, only the dimensions are compared. Shape may be compared in the future for tighter type inference. Satisfies the hm.Type interface.

func (TensorType) Format ¶ Uses

func (t TensorType) Format(state fmt.State, c rune)

Format implements fmt.Formatter. It is also required for the satisfication the hm.Type interface.

func (TensorType) FreeTypeVar ¶ Uses

func (t TensorType) FreeTypeVar() hm.TypeVarSet

FreeTypeVar returns any free (unbound) type variables in this type. Satisfies the hm.Type interface.

func (TensorType) Name ¶ Uses

func (t TensorType) Name() string

Name returns the name of the type, which will always be "Tensor". Satisfies the hm.Type interface.

func (TensorType) Normalize ¶ Uses

func (t TensorType) Normalize(k, v hm.TypeVarSet) (hm.Type, error)

Normalize normalizes the type variable names (if any) in the TensorType. Satisfies the hm.Type interface.

func (TensorType) String ¶ Uses

func (t TensorType) String() string

String implements fmt.Stringer and runtime.Stringer. Satisfies the hm.Type interface.

func (TensorType) Types ¶ Uses

func (t TensorType) Types() hm.Types

Types returns a list of types that TensorType contains - in this case, the type of Tensor (float64, float32, etc). Satisfies the hm.Type interface.

type Typer ¶ Uses

type Typer interface {
    Type() hm.Type
}

Typer represents any type (typically a Op) that knows its own Type

type U8 ¶ Uses

type U8 byte

U8 represents a byte value.

func (*U8) Data ¶ Uses

func (v *U8) Data() interface{}

Data returns the original representation of the Value

func (*U8) Dtype ¶ Uses

func (v *U8) Dtype() tensor.Dtype

Dtype returns the Dtype of the value

func (*U8) Format ¶ Uses

func (v *U8) Format(s fmt.State, c rune)

Format implements fmt.Formatter

func (*U8) MemSize ¶ Uses

func (v *U8) MemSize() uintptr

MemSize satisfies the tensor.Memory interface

func (*U8) Pointer ¶ Uses

func (v *U8) Pointer() unsafe.Pointer

Pointer returns the pointer as an unsafe.Pointer. Satisfies the tensor.Memory interface

func (*U8) Shape ¶ Uses

func (v *U8) Shape() tensor.Shape

Shape returns a scalar shape for all scalar values

func (*U8) Size ¶ Uses

func (v *U8) Size() int

Size returns 0 for all scalar Values

func (*U8) Uintptr ¶ Uses

func (v *U8) Uintptr() uintptr

Uintptr satisfies the tensor.Memory interface

type UnaryOp ¶ Uses

type UnaryOp interface {
    Op

    IsUnary() bool
}

A UnaryOp is an Op that takes only one input

type UnsafeDoer ¶ Uses

type UnsafeDoer interface {
    UnsafeDo(inputs ...Value) (Value, error)
}

UnsafeDoer is an op that will overwrite the underlying value.

type UsePreallocDoer ¶ Uses

type UsePreallocDoer interface {
    UsePreallocDo(prealloc Value, inputs ...Value) (Value, error)
}

UsePreallocDoer is an op that works when a preallocated value is provided

type VM ¶ Uses

type VM interface {
    RunAll() error
    Reset()

    // Close closes all the machine resources (CUDA, if any, loggers if any)
    Close() error
}

VM represents a structure that can execute a graph or program. There are two VMs (both unexported):

- *tapeMachine
- *lispMachine

The *tapeMachine pre-compiles a graph into a list of instructions, then executes the instructions linearly and sequentially. The main tradeoff is dynamism. Graphs cannot be dynamically created on the fly as a re-compilation process is required (and compilation is relatively expensive). However, graphs executed with the *tapeMachine run much faster as plenty of optimizations has been done in the code generation stage.

The *lispMachine allows for graphs to be dynamically built and executed upon. The tradeoff is that executing a graph on *lispMachine is generally slower than on *tapeMachine, given the same static "image" of a graph.

type VMOpt ¶ Uses

type VMOpt func(m VM)

VMOpt is a VM creation option

func BindDualValues ¶ Uses

func BindDualValues(nodes ...*Node) VMOpt

BindDualValues is an option for *tapeMachine only. This is useful to set when using a Solver

func ExecuteBwdOnly ¶ Uses

func ExecuteBwdOnly() VMOpt

ExecuteBwdOnly creates a VM that will execute a graph by doing back propagation only. The assumption is of course, that the forward graph has already been executed, and there are already values associated with the nodes. This option is only for *lispMachine. Try it on any other VMs and it will panic.

func ExecuteFwdOnly ¶ Uses

func ExecuteFwdOnly() VMOpt

ExecuteFwdOnly creates a VM that will execute a graph forwards only - it will not do back propagation. This option is only for *lispMachine. Try it on any other VMs and it will panic.

func LogBothDir ¶ Uses

func LogBothDir() VMOpt

LogBothDir logs both directions of the execution of the graph. This option is only available for *lispMachine.

func LogBwd ¶ Uses

func LogBwd() VMOpt

LogBwd logs the backwards execution of a graph. This option is only for *lispMachine. Try it on any other VMs and it will panic.

func LogFwd ¶ Uses

func LogFwd() VMOpt

LogFwd logs the forward execution of a graph. This option is only for *lispMachine. Try it on any other VMs and it will panic.

func TraceExec ¶ Uses

func TraceExec() VMOpt

TraceExec is an option for *tapeMachine only. It stores an immutable copy of the executed value into the node, instead of a mutable value, which may be clobbered

func UseCudaFor ¶ Uses

func UseCudaFor(ops ...string) VMOpt

UseCudaFor is an option for *tapeMachine. This function is NO-OP unless the program is built with the `cuda` tag.

func WithEngine ¶ Uses

func WithEngine(e tensor.Engine) VMOpt

WithEngine sets the tensor engine for computation inside the VM.

func WithInfWatch ¶ Uses

func WithInfWatch() VMOpt

WithInfWatch creates a VM that will watch for Infs when executing. It watches for +Inf, -Inf and Inf. No choice there. This slows the execution down.

func WithLogger ¶ Uses

func WithLogger(logger *log.Logger) VMOpt

WithLogger creates a VM with the supplied logger. If the logger is nil, a default logger, writing to os.stderr will be created.

func WithManualGradient ¶ Uses

func WithManualGradient() VMOpt

WithManualGradient allows the user to set the gradient of the root, before backprop. The root gradients should be set using the SetDeriv method

func WithNaNWatch ¶ Uses

func WithNaNWatch() VMOpt

WithNaNWatch creates a VM that will watch for NaNs when executing. This slows the execution down.

func WithPrecompiled ¶ Uses

func WithPrecompiled(prog *program, locMap map[*Node]register) VMOpt

WithPrecompiled is an option to pass in compiled programs. This is useful for users who use the CompileFunction function

func WithValueFmt ¶ Uses

func WithValueFmt(format string) VMOpt

WithValueFmt defines how the logger will output the values. It defaults to "%3.3f"

func WithWatchlist ¶ Uses

func WithWatchlist(list ...interface{}) VMOpt

WithWatchlist creates a VM with a watchlist. When the execution touches the things in the watchlist, the VM's logger will the log it. This allows for watching and finetuning of the algorithm. When nothing is passed in, then the VM will default to watching and logging every single execution object.

The watchlist allows for different things to be watched, depending on VM type:

*lispMachine will ONLY take *Node
*tapeMachine will take int (for register IDs) or *Node.

type Value ¶ Uses

type Value interface {
    Shape() tensor.Shape // Shape  returns the shape of the Value. Scalar values return ScalarShape()
    Size() int           // Size represents the number of elements in the Value. Note that in cases such as a *tensor.Dense, the underlying slice MAY have more elements than the Size() reports. This is correct.
    Data() interface{}   // Data returns the original representation of the Value
    Dtype() tensor.Dtype // Dtype returns the Dtype of the value

    tensor.Memory
    fmt.Formatter
}

Value represents a value that Gorgonia accepts. At this point it is implemented by:

- all scalar value types (F64, F32... etc)
- *tensor.Dense
- *dualValue

A Value is essentially any thing that knows its own type and shape. Most importantly though, a Value is a pointer - and can be converted into a tensor.Memory. This is done for the sake of interoperability with external devices like cgo or CUDA or OpenCL. This also means for the most part most Values will be allocated on the heap. There are some performance tradeoffs made in this decision, but ultimately this is better than having to manually manage blocks of memory

func CloneValue ¶ Uses

func CloneValue(v Value) (Value, error)

CloneValue clones a value. For scalars, since Go copies scalars, it returns itself

func Copy ¶ Uses

func Copy(dest, src Value) (Value, error)

Copy copies the src values into dest values. For scalars, it just returns itself

func ScalarAsTensor ¶ Uses

func ScalarAsTensor(v Value, dims int, e tensor.Engine) Value

ScalarAsTensor returns the tensor representation of a scalar. It is particularly useful as a "reshape" of tensors of sorts

The Value passed in are either Scalar, tensor.Tensor, or *dualValue. Anything else will panic.

func ZeroValue ¶ Uses

func ZeroValue(v Value) Value

ZeroValue returns the zero value of a type

type ValueCloser ¶ Uses

type ValueCloser interface {
    ValueClose(interface{}) bool
}

ValueCloser represents any type that can perform a close-value check

type ValueEqualer ¶ Uses

type ValueEqualer interface {
    ValueEq(Value) bool
}

ValueEqualer represents any type that can perform a equal value check

type ValueGrad ¶ Uses

type ValueGrad interface {
    Valuer
    Grad() (Value, error)
}

ValueGrad is any type that has a value and a grad. This is used for Solvers

func NodesToValueGrads ¶ Uses

func NodesToValueGrads(in Nodes) (out []ValueGrad)

NodesToValueGrads is a utility function that converts a Nodes to a slice of ValueGrad for the solvers

type Valuer ¶ Uses

type Valuer interface {
    Value() Value
}

Valuer is any type that can return a Value

type VanillaSolver ¶ Uses

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

VanillaSolver is your bog standard stochastic gradient descent optimizer. There are no fancy features to this

func NewVanillaSolver ¶ Uses

func NewVanillaSolver(opts ...SolverOpt) *VanillaSolver

NewVanillaSolver creates a new VanillaSolver with sane-ish default values

func (*VanillaSolver) Step ¶ Uses

func (s *VanillaSolver) Step(model []ValueGrad) (err error)

Step steps through each node in the model and applies the most basic gradient descent algorithm on the value.

This function will error out if the nodes do not have an associated Grad value.

type ZeroValuer ¶ Uses

type ZeroValuer interface {
    Value
    ZeroValue() Value
}

ZeroValuer is a a Value that can provide the zero-value of its type

type Zeroer ¶ Uses

type Zeroer interface {
    Value
    Zero()
}

Zeroer is a Value that can zero itself

Directories ¶