Feature/diff api #154

Merged
altavir merged 20 commits from feature/diff-api into dev 2020-10-28 13:25:24 +03:00
5 changed files with 190 additions and 1 deletions
Showing only changes of commit d826dd9e83 - Show all commits

View File

@ -8,6 +8,9 @@
- Automatic README generation for features (#139) - Automatic README generation for features (#139)
- Native support for `memory`, `core` and `dimensions` - Native support for `memory`, `core` and `dimensions`
- `kmath-ejml` to supply EJML SimpleMatrix wrapper. - `kmath-ejml` to supply EJML SimpleMatrix wrapper.
- A separate `Symbol` entity, which is used for global unbound symbol.
- A `Symbol` indexing scope.
- Basic optimization API for Commons-math.
### Changed ### Changed
- Package changed from `scientifik` to `kscience.kmath`. - Package changed from `scientifik` to `kscience.kmath`.
@ -16,6 +19,7 @@
- `Polynomial` secondary constructor made function. - `Polynomial` secondary constructor made function.
- Kotlin version: 1.3.72 -> 1.4.20-M1 - Kotlin version: 1.3.72 -> 1.4.20-M1
- `kmath-ast` doesn't depend on heavy `kotlin-reflect` library. - `kmath-ast` doesn't depend on heavy `kotlin-reflect` library.
- Full autodiff refactoring based on `Symbol`
### Deprecated ### Deprecated

View File

@ -0,0 +1,103 @@
package kscience.kmath.commons.optimization
import kscience.kmath.expressions.*
import org.apache.commons.math3.optim.*
import org.apache.commons.math3.optim.nonlinear.scalar.GoalType
import org.apache.commons.math3.optim.nonlinear.scalar.MultivariateOptimizer
import org.apache.commons.math3.optim.nonlinear.scalar.ObjectiveFunction
import org.apache.commons.math3.optim.nonlinear.scalar.ObjectiveFunctionGradient
import org.apache.commons.math3.optim.nonlinear.scalar.gradient.NonLinearConjugateGradientOptimizer
import org.apache.commons.math3.optim.nonlinear.scalar.noderiv.NelderMeadSimplex
import org.apache.commons.math3.optim.nonlinear.scalar.noderiv.SimplexOptimizer
public typealias ParameterSpacePoint = Map<Symbol, Double>
public class OptimizationResult(public val point: ParameterSpacePoint, public val value: Double)
public operator fun PointValuePair.component1(): DoubleArray = point
public operator fun PointValuePair.component2(): Double = value
public object Optimization {
public const val DEFAULT_RELATIVE_TOLERANCE: Double = 1e-4
public const val DEFAULT_ABSOLUTE_TOLERANCE: Double = 1e-4
public const val DEFAULT_MAX_ITER: Int = 1000
}
private fun SymbolIndexer.objectiveFunction(expression: Expression<Double>) = ObjectiveFunction {
val args = it.toMap()
expression(args)
}
private fun SymbolIndexer.objectiveFunctionGradient(
expression: DifferentiableExpression<Double>,
) = ObjectiveFunctionGradient {
val args = it.toMap()
DoubleArray(symbols.size) { index ->
expression.derivative(symbols[index])(args)
}
}
private fun SymbolIndexer.initialGuess(point: ParameterSpacePoint) = InitialGuess(point.toArray())
/**
* Optimize expression without derivatives
*/
public fun Expression<Double>.optimize(
startingPoint: ParameterSpacePoint,
goalType: GoalType = GoalType.MAXIMIZE,
vararg additionalArguments: OptimizationData,
optimizerBuilder: () -> MultivariateOptimizer = {
SimplexOptimizer(
SimpleValueChecker(
Optimization.DEFAULT_RELATIVE_TOLERANCE,
Optimization.DEFAULT_ABSOLUTE_TOLERANCE,
Optimization.DEFAULT_MAX_ITER
)
)
},
): OptimizationResult = withSymbols(startingPoint.keys) {
val optimizer = optimizerBuilder()
val objectiveFunction = objectiveFunction(this@optimize)
val (point, value) = optimizer.optimize(
objectiveFunction,
initialGuess(startingPoint),
goalType,
MaxEval.unlimited(),
NelderMeadSimplex(symbols.size, 1.0),
*additionalArguments
)
OptimizationResult(point.toMap(), value)
}
/**
* Optimize differentiable expression
*/
public fun DifferentiableExpression<Double>.optimize(
startingPoint: ParameterSpacePoint,
goalType: GoalType = GoalType.MAXIMIZE,
vararg additionalArguments: OptimizationData,
optimizerBuilder: () -> NonLinearConjugateGradientOptimizer = {
NonLinearConjugateGradientOptimizer(
NonLinearConjugateGradientOptimizer.Formula.FLETCHER_REEVES,
SimpleValueChecker(
Optimization.DEFAULT_RELATIVE_TOLERANCE,
Optimization.DEFAULT_ABSOLUTE_TOLERANCE,
Optimization.DEFAULT_MAX_ITER
)
)
},
): OptimizationResult = withSymbols(startingPoint.keys) {
val optimizer = optimizerBuilder()
val objectiveFunction = objectiveFunction(this@optimize)
val objectiveGradient = objectiveFunctionGradient(this@optimize)
val (point, value) = optimizer.optimize(
objectiveFunction,
objectiveGradient,
initialGuess(startingPoint),
goalType,
MaxEval.unlimited(),
*additionalArguments
)
OptimizationResult(point.toMap(), value)
}

View File

@ -0,0 +1,37 @@
package kscience.kmath.commons.optimization
import kscience.kmath.commons.expressions.DerivativeStructureExpression
import kscience.kmath.expressions.Expression
import kscience.kmath.expressions.Symbol
import kscience.kmath.expressions.symbol
import org.apache.commons.math3.optim.nonlinear.scalar.noderiv.SimplexOptimizer
import org.junit.jupiter.api.Test
internal class OptimizeTest {
val x by symbol
val y by symbol
val normal = DerivativeStructureExpression {
val x = bind(x)
val y = bind(y)
exp(-x.pow(2)/2) + exp(-y.pow(2)/2)
}
val startingPoint: Map<Symbol, Double> = mapOf(x to 1.0, y to 1.0)
@Test
fun testOptimization() {
val result = normal.optimize(startingPoint)
println(result.point)
println(result.value)
}
@Test
fun testSimplexOptimization() {
val result = (normal as Expression<Double>).optimize(startingPoint){
SimplexOptimizer(1e-4,1e-4)
}
println(result.point)
println(result.value)
}
}

View File

@ -35,7 +35,7 @@ public fun interface Expression<T> {
} }
/** /**
* Invlode an expression without parameters * Invoke an expression without parameters
*/ */
public operator fun <T> Expression<T>.invoke(): T = invoke(emptyMap()) public operator fun <T> Expression<T>.invoke(): T = invoke(emptyMap())
//This method exists to avoid resolution ambiguity of vararg methods //This method exists to avoid resolution ambiguity of vararg methods

View File

@ -0,0 +1,45 @@
package kscience.kmath.expressions
/**
* An environment to easy transform indexed variables to symbols and back.
*/
public interface SymbolIndexer {
public val symbols: List<Symbol>
public fun indexOf(symbol: Symbol): Int = symbols.indexOf(symbol)
public operator fun <T> List<T>.get(symbol: Symbol): T {
require(size == symbols.size) { "The input list size for indexer should be ${symbols.size} but $size found" }
return get(this@SymbolIndexer.indexOf(symbol))
}
public operator fun <T> Array<T>.get(symbol: Symbol): T {
require(size == symbols.size) { "The input array size for indexer should be ${symbols.size} but $size found" }
return get(this@SymbolIndexer.indexOf(symbol))
}
public operator fun DoubleArray.get(symbol: Symbol): Double {
require(size == symbols.size) { "The input array size for indexer should be ${symbols.size} but $size found" }
return get(this@SymbolIndexer.indexOf(symbol))
}
public fun DoubleArray.toMap(): Map<Symbol, Double> {
require(size == symbols.size) { "The input array size for indexer should be ${symbols.size} but $size found" }
return symbols.indices.associate { symbols[it] to get(it) }
}
public fun <T> Map<Symbol, T>.toList(): List<T> = symbols.map { getValue(it) }
public fun Map<Symbol, Double>.toArray(): DoubleArray = DoubleArray(symbols.size) { getValue(symbols[it]) }
}
public inline class SimpleSymbolIndexer(override val symbols: List<Symbol>) : SymbolIndexer
/**
* Execute the block with symbol indexer based on given symbol order
*/
public inline fun <R> withSymbols(vararg symbols: Symbol, block: SymbolIndexer.() -> R): R =
with(SimpleSymbolIndexer(symbols.toList()), block)
public inline fun <R> withSymbols(symbols: Collection<Symbol>, block: SymbolIndexer.() -> R): R =
with(SimpleSymbolIndexer(symbols.toList()), block)