Safe modification of autodiff

This commit is contained in:
Alexander Nozik 2019-05-05 10:14:05 +03:00
parent 6f9b704aa7
commit c3f0dbe161
3 changed files with 93 additions and 47 deletions

View File

@ -1,5 +1,7 @@
package scientifik.kmath.operations
import scientifik.kmath.linear.Point
import scientifik.kmath.structures.asBuffer
import kotlin.math.pow
import kotlin.math.sqrt
@ -12,12 +14,26 @@ import kotlin.math.sqrt
* Differentiable variable with value and derivative of differentiation ([deriv]) result
* with respect to this variable.
*/
open class Variable(val x: Double) {
open class Variable(val value: Double) {
constructor(x: Number) : this(x.toDouble())
}
class DerivationResult(x: Double, val deriv: Map<Variable, Double>): Variable(x) {
class DerivationResult(value: Double, val deriv: Map<Variable, Double>) : Variable(value) {
fun deriv(variable: Variable) = deriv[variable] ?: 0.0
/**
* compute divergence
*/
fun div() = deriv.values.sum()
/**
* Compute a gradient for variables in given order
*/
fun grad(vararg variables: Variable): Point<Double> = if (variables.isEmpty()) {
error("Variable order is not provided for gradient construction")
} else {
variables.map(::deriv).toDoubleArray().asBuffer()
}
}
/**
@ -38,7 +54,7 @@ fun deriv(body: AutoDiffField.() -> Variable): DerivationResult =
val result = body()
result.d = 1.0 // computing derivative w.r.t result
runBackwardPass()
DerivationResult(result.x, derivatives)
DerivationResult(result.value, derivatives)
}
@ -61,19 +77,21 @@ abstract class AutoDiffField : Field<Variable> {
*/
abstract var Variable.d: Double
abstract fun variable(value: Double): Variable
// Overloads for Double constants
operator fun Number.plus(that: Variable): Variable = derive(Variable(this.toDouble() + that.x)) { z ->
operator fun Number.plus(that: Variable): Variable = derive(variable(this.toDouble() + that.value)) { z ->
that.d += z.d
}
operator fun Variable.plus(b: Number): Variable = b.plus(this)
operator fun Number.minus(that: Variable): Variable = derive(Variable(this.toDouble() - that.x)) { z ->
operator fun Number.minus(that: Variable): Variable = derive(variable(this.toDouble() - that.value)) { z ->
that.d -= z.d
}
operator fun Variable.minus(that: Number): Variable = derive(Variable(this.x - that.toDouble())) { z ->
operator fun Variable.minus(that: Number): Variable = derive(variable(this.value - that.toDouble())) { z ->
this.d += z.d
}
}
@ -89,11 +107,23 @@ private class AutoDiffContext : AutoDiffField() {
internal val derivatives = HashMap<Variable, Double>()
/**
* A variable coupled with its derivative. For internal use only
*/
class VariableWithDeriv(x: Double, var d: Double = 0.0): Variable(x)
override fun variable(value: Double): Variable = VariableWithDeriv(value)
override var Variable.d: Double
get() = derivatives[this] ?: 0.0
get() = (this as? VariableWithDeriv)?.d ?: derivatives[this] ?: 0.0
set(value) {
if(this is VariableWithDeriv){
d = value
}else {
derivatives[this] = value
}
}
@Suppress("UNCHECKED_CAST")
override fun <R> derive(value: R, block: (R) -> Unit): R {
@ -117,25 +147,25 @@ private class AutoDiffContext : AutoDiffField() {
override fun add(a: Variable, b: Variable): Variable =
derive(Variable(a.x + b.x)) { z ->
derive(variable(a.value + b.value)) { z ->
a.d += z.d
b.d += z.d
}
override fun multiply(a: Variable, b: Variable): Variable =
derive(Variable(a.x * b.x)) { z ->
a.d += z.d * b.x
b.d += z.d * a.x
derive(variable(a.value * b.value)) { z ->
a.d += z.d * b.value
b.d += z.d * a.value
}
override fun divide(a: Variable, b: Variable): Variable =
derive(Variable(a.x / b.x)) { z ->
a.d += z.d / b.x
b.d -= z.d * a.x / (b.x * b.x)
derive(Variable(a.value / b.value)) { z ->
a.d += z.d / b.value
b.d -= z.d * a.value / (b.value * b.value)
}
override fun multiply(a: Variable, k: Number): Variable =
derive(Variable(k.toDouble() * a.x)) { z ->
derive(variable(k.toDouble() * a.value)) { z ->
a.d += z.d * k.toDouble()
}
@ -146,41 +176,41 @@ private class AutoDiffContext : AutoDiffField() {
// Extensions for differentiation of various basic mathematical functions
// x ^ 2
fun AutoDiffField.sqr(x: Variable): Variable = derive(Variable(x.x * x.x)) { z ->
x.d += z.d * 2 * x.x
fun AutoDiffField.sqr(x: Variable): Variable = derive(variable(x.value * x.value)) { z ->
x.d += z.d * 2 * x.value
}
// x ^ 1/2
fun AutoDiffField.sqrt(x: Variable): Variable = derive(Variable(sqrt(x.x))) { z ->
x.d += z.d * 0.5 / z.x
fun AutoDiffField.sqrt(x: Variable): Variable = derive(variable(sqrt(x.value))) { z ->
x.d += z.d * 0.5 / z.value
}
// x ^ y (const)
fun AutoDiffField.pow(x: Variable, y: Double): Variable = derive(Variable(x.x.pow(y))) { z ->
x.d += z.d * y * x.x.pow(y - 1)
fun AutoDiffField.pow(x: Variable, y: Double): Variable = derive(variable(x.value.pow(y))) { z ->
x.d += z.d * y * x.value.pow(y - 1)
}
fun AutoDiffField.pow(x: Variable, y: Int): Variable = pow(x, y.toDouble())
// exp(x)
fun AutoDiffField.exp(x: Variable): Variable = derive(Variable(kotlin.math.exp(x.x))) { z ->
x.d += z.d * z.x
fun AutoDiffField.exp(x: Variable): Variable = derive(variable(kotlin.math.exp(x.value))) { z ->
x.d += z.d * z.value
}
// ln(x)
fun AutoDiffField.ln(x: Variable): Variable = derive(Variable(kotlin.math.ln(x.x))) { z ->
x.d += z.d / x.x
fun AutoDiffField.ln(x: Variable): Variable = derive(Variable(kotlin.math.ln(x.value))) { z ->
x.d += z.d / x.value
}
// x ^ y (any)
fun AutoDiffField.pow(x: Variable, y: Variable): Variable = exp(y * ln(x))
// sin(x)
fun AutoDiffField.sin(x: Variable): Variable = derive(Variable(kotlin.math.sin(x.x))) { z ->
x.d += z.d * kotlin.math.cos(x.x)
fun AutoDiffField.sin(x: Variable): Variable = derive(variable(kotlin.math.sin(x.value))) { z ->
x.d += z.d * kotlin.math.cos(x.value)
}
// cos(x)
fun AutoDiffField.cos(x: Variable): Variable = derive(Variable(kotlin.math.cos(x.x))) { z ->
x.d -= z.d * kotlin.math.sin(x.x)
fun AutoDiffField.cos(x: Variable): Variable = derive(variable(kotlin.math.cos(x.value))) { z ->
x.d -= z.d * kotlin.math.sin(x.value)
}

View File

@ -37,6 +37,11 @@ interface Buffer<T> {
companion object {
inline fun real(size: Int, initializer: (Int) -> Double): DoubleBuffer {
val array = DoubleArray(size) { initializer(it) }
return DoubleBuffer(array)
}
/**
* Create a boxing buffer of given type
*/

View File

@ -1,15 +1,17 @@
package scientifik.kmath.operations
import scientifik.kmath.structures.asBuffer
import kotlin.math.PI
import kotlin.test.Test
import kotlin.test.assertEquals
import kotlin.test.assertTrue
class AutoDiffTest {
@Test
fun testPlusX2() {
val x = Variable(3) // diff w.r.t this x at 3
val y = deriv { x + x }
assertEquals(6.0, y.x) // y = x + x = 6
assertEquals(6.0, y.value) // y = x + x = 6
assertEquals(2.0, y.deriv(x)) // dy/dx = 2
}
@ -19,7 +21,7 @@ class AutoDiffTest {
val x = Variable(2)
val y = Variable(3)
val z = deriv { x + y }
assertEquals(5.0, z.x) // z = x + y = 5
assertEquals(5.0, z.value) // z = x + y = 5
assertEquals(1.0, z.deriv(x)) // dz/dx = 1
assertEquals(1.0, z.deriv(y)) // dz/dy = 1
}
@ -30,7 +32,7 @@ class AutoDiffTest {
val x = Variable(7)
val y = Variable(3)
val z = deriv { x - y }
assertEquals(4.0, z.x) // z = x - y = 4
assertEquals(4.0, z.value) // z = x - y = 4
assertEquals(1.0, z.deriv(x)) // dz/dx = 1
assertEquals(-1.0, z.deriv(y)) // dz/dy = -1
}
@ -39,7 +41,7 @@ class AutoDiffTest {
fun testMulX2() {
val x = Variable(3) // diff w.r.t this x at 3
val y = deriv { x * x }
assertEquals(9.0, y.x) // y = x * x = 9
assertEquals(9.0, y.value) // y = x * x = 9
assertEquals(6.0, y.deriv(x)) // dy/dx = 2 * x = 7
}
@ -47,7 +49,7 @@ class AutoDiffTest {
fun testSqr() {
val x = Variable(3)
val y = deriv { sqr(x) }
assertEquals(9.0, y.x) // y = x ^ 2 = 9
assertEquals(9.0, y.value) // y = x ^ 2 = 9
assertEquals(6.0, y.deriv(x)) // dy/dx = 2 * x = 7
}
@ -55,7 +57,7 @@ class AutoDiffTest {
fun testSqrSqr() {
val x = Variable(2)
val y = deriv { sqr(sqr(x)) }
assertEquals(16.0, y.x) // y = x ^ 4 = 16
assertEquals(16.0, y.value) // y = x ^ 4 = 16
assertEquals(32.0, y.deriv(x)) // dy/dx = 4 * x^3 = 32
}
@ -63,7 +65,7 @@ class AutoDiffTest {
fun testX3() {
val x = Variable(2) // diff w.r.t this x at 2
val y = deriv { x * x * x }
assertEquals(8.0, y.x) // y = x * x * x = 8
assertEquals(8.0, y.value) // y = x * x * x = 8
assertEquals(12.0, y.deriv(x)) // dy/dx = 3 * x * x = 12
}
@ -72,7 +74,7 @@ class AutoDiffTest {
val x = Variable(5)
val y = Variable(2)
val z = deriv { x / y }
assertEquals(2.5, z.x) // z = x / y = 2.5
assertEquals(2.5, z.value) // z = x / y = 2.5
assertEquals(0.5, z.deriv(x)) // dz/dx = 1 / y = 0.5
assertEquals(-1.25, z.deriv(y)) // dz/dy = -x / y^2 = -1.25
}
@ -81,7 +83,7 @@ class AutoDiffTest {
fun testPow3() {
val x = Variable(2) // diff w.r.t this x at 2
val y = deriv { pow(x, 3) }
assertEquals(8.0, y.x) // y = x ^ 3 = 8
assertEquals(8.0, y.value) // y = x ^ 3 = 8
assertEquals(12.0, y.deriv(x)) // dy/dx = 3 * x ^ 2 = 12
}
@ -90,7 +92,7 @@ class AutoDiffTest {
val x = Variable(2)
val y = Variable(3)
val z = deriv { pow(x, y) }
assertApprox(8.0, z.x) // z = x ^ y = 8
assertApprox(8.0, z.value) // z = x ^ y = 8
assertApprox(12.0, z.deriv(x)) // dz/dx = y * x ^ (y - 1) = 12
assertApprox(8.0 * kotlin.math.ln(2.0), z.deriv(y)) // dz/dy = x ^ y * ln(x)
}
@ -99,7 +101,7 @@ class AutoDiffTest {
fun testFromPaper() {
val x = Variable(3)
val y = deriv { 2 * x + x * x * x }
assertEquals(33.0, y.x) // y = 2 * x + x * x * x = 33
assertEquals(33.0, y.value) // y = 2 * x + x * x * x = 33
assertEquals(29.0, y.deriv(x)) // dy/dx = 2 + 3 * x * x = 29
}
@ -109,7 +111,7 @@ class AutoDiffTest {
val y = deriv {
Variable(1) * x
}
assertEquals(1.0, y.x) // y = x ^ n = 1
assertEquals(1.0, y.value) // y = x ^ n = 1
assertEquals(1.0, y.deriv(x)) // dy/dx = n * x ^ (n - 1) = n - 1
}
@ -122,7 +124,7 @@ class AutoDiffTest {
for (i in 1..n) res *= x
res
}
assertEquals(1.0, y.x) // y = x ^ n = 1
assertEquals(1.0, y.value) // y = x ^ n = 1
assertEquals(n.toDouble(), y.deriv(x)) // dy/dx = n * x ^ (n - 1) = n - 1
}
@ -130,7 +132,7 @@ class AutoDiffTest {
fun testExample() {
val x = Variable(2)
val y = deriv { sqr(x) + 5 * x + 3 }
assertEquals(17.0, y.x) // the value of result (y)
assertEquals(17.0, y.value) // the value of result (y)
assertEquals(9.0, y.deriv(x)) // dy/dx
}
@ -138,7 +140,7 @@ class AutoDiffTest {
fun testSqrt() {
val x = Variable(16)
val y = deriv { sqrt(x) }
assertEquals(4.0, y.x) // y = x ^ 1/2 = 4
assertEquals(4.0, y.value) // y = x ^ 1/2 = 4
assertEquals(1.0 / 8, y.deriv(x)) // dy/dx = 1/2 / x ^ 1/4 = 1/8
}
@ -146,7 +148,7 @@ class AutoDiffTest {
fun testSin() {
val x = Variable(PI / 6)
val y = deriv { sin(x) }
assertApprox(0.5, y.x) // y = sin(PI/6) = 0.5
assertApprox(0.5, y.value) // y = sin(PI/6) = 0.5
assertApprox(kotlin.math.sqrt(3.0) / 2, y.deriv(x)) // dy/dx = cos(PI/6) = sqrt(3)/2
}
@ -154,10 +156,19 @@ class AutoDiffTest {
fun testCos() {
val x = Variable(PI / 6)
val y = deriv { cos(x) }
assertApprox(kotlin.math.sqrt(3.0) / 2, y.x) // y = cos(PI/6) = sqrt(3)/2
assertApprox(kotlin.math.sqrt(3.0) / 2, y.value) // y = cos(PI/6) = sqrt(3)/2
assertApprox(-0.5, y.deriv(x)) // dy/dx = -sin(PI/6) = -0.5
}
@Test
fun testDivGrad() {
val x = Variable(1.0)
val y = Variable(2.0)
val res = deriv { x * x + y * y }
assertEquals(6.0, res.div())
assertTrue(res.grad(x, y).contentEquals(doubleArrayOf(2.0, 4.0).asBuffer()))
}
private fun assertApprox(a: Double, b: Double) {
if ((a - b) > 1e-10) assertEquals(a, b)
}