Files
QuanTAlib/lib/errors/huber/Huber.cs
T
Miha Kralj 6e24fea8b7 Add Tukey's Biweight and WMAPE implementations with comprehensive tests and documentation
- Introduced Tukey's Biweight as a robust loss function, including mathematical foundation, usage patterns, and performance profile.
- Added WMAPE (Weighted Mean Absolute Percentage Error) implementation, emphasizing its advantages for intermittent demand forecasting.
- Created unit tests for WMAPE covering various scenarios including edge cases and batch calculations.
- Documented both Tukey's Biweight and WMAPE with detailed explanations, properties, and common use cases.
2025-12-30 09:27:08 -08:00

361 lines
13 KiB
C#

using System.Numerics;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.X86;
namespace QuanTAlib;
/// <summary>
/// Huber: Huber Loss
/// </summary>
/// <remarks>
/// Huber Loss combines the best properties of MSE and MAE. For small errors
/// (|error| ≤ delta), it behaves like MSE (quadratic). For large errors
/// (|error| > delta), it behaves like MAE (linear).
///
/// Formula:
/// If |error| ≤ delta: L = 0.5 * error²
/// If |error| > delta: L = delta * |error| - 0.5 * delta²
///
/// Key properties:
/// - Differentiable everywhere (unlike MAE)
/// - Robust to outliers (unlike MSE)
/// - Delta controls the transition point
/// - Default delta = 1.345 (for 95% efficiency with normal distribution)
/// </remarks>
[SkipLocalsInit]
public sealed class Huber : AbstractBase
{
private readonly double _delta;
private readonly double _halfDeltaSquared;
private readonly RingBuffer _buffer;
[StructLayout(LayoutKind.Auto)]
private record struct State(double Sum, double LastValidActual, double LastValidPredicted, int TickCount);
private State _state;
private State _p_state;
private const int ResyncInterval = 1000;
public Huber(int period, double delta = 1.345)
{
if (period <= 0)
throw new ArgumentException("Period must be greater than 0", nameof(period));
if (delta <= 0)
throw new ArgumentException("Delta must be greater than 0", nameof(delta));
_delta = delta;
_halfDeltaSquared = 0.5 * delta * delta;
_buffer = new RingBuffer(period);
Name = $"Huber({period},{delta:F3})";
WarmupPeriod = period;
}
public override bool IsHot => _buffer.IsFull;
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private double CalculateHuberLoss(double error)
{
double absError = Math.Abs(error);
// Use FMA for the linear portion: delta * absError - halfDeltaSquared
// = FMA(delta, absError, -halfDeltaSquared)
return absError <= _delta
? 0.5 * error * error
: Math.FusedMultiplyAdd(_delta, absError, -_halfDeltaSquared);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public TValue Update(TValue actual, TValue predicted, bool isNew = true)
{
double actualVal = actual.Value;
double predictedVal = predicted.Value;
if (!double.IsFinite(actualVal))
actualVal = double.IsFinite(_state.LastValidActual) ? _state.LastValidActual : 0.0;
else
_state.LastValidActual = actualVal;
if (!double.IsFinite(predictedVal))
predictedVal = double.IsFinite(_state.LastValidPredicted) ? _state.LastValidPredicted : 0.0;
else
_state.LastValidPredicted = predictedVal;
double error = actualVal - predictedVal;
double huberLoss = CalculateHuberLoss(error);
if (isNew)
{
_p_state = _state;
double removedValue = _buffer.Count == _buffer.Capacity ? _buffer.Oldest : 0.0;
_state.Sum = _state.Sum - removedValue + huberLoss;
_buffer.Add(huberLoss);
_state.TickCount++;
if (_buffer.IsFull && _state.TickCount >= ResyncInterval)
{
_state.TickCount = 0;
_state.Sum = _buffer.RecalculateSum();
}
}
else
{
_state = _p_state;
double removedValue = _buffer.Count == _buffer.Capacity ? _buffer.Oldest : 0.0;
_state.Sum = _state.Sum - removedValue + huberLoss;
_buffer.UpdateNewest(huberLoss);
_state.Sum = _buffer.RecalculateSum();
}
double result = _buffer.Count > 0 ? _state.Sum / _buffer.Count : huberLoss;
Last = new TValue(actual.Time, result);
PubEvent(Last, isNew);
return Last;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public TValue Update(double actual, double predicted, bool isNew = true)
{
return Update(new TValue(DateTime.UtcNow, actual), new TValue(DateTime.UtcNow, predicted), isNew);
}
public override TValue Update(TValue input, bool isNew = true)
{
throw new NotSupportedException("Huber requires two inputs. Use Update(actual, predicted).");
}
public override TSeries Update(TSeries source)
{
throw new NotSupportedException("Huber requires two inputs. Use Calculate(actualSeries, predictedSeries, period, delta).");
}
public override void Prime(ReadOnlySpan<double> source, TimeSpan? step = null)
{
throw new NotSupportedException("Huber requires two inputs.");
}
public override void Reset()
{
_buffer.Clear();
_state = default;
_p_state = default;
Last = default;
}
public static TSeries Calculate(TSeries actual, TSeries predicted, int period, double delta = 1.345)
{
if (actual.Count != predicted.Count)
throw new ArgumentException("Actual and predicted series must have the same length", nameof(predicted));
int len = actual.Count;
var t = new List<long>(len);
var v = new List<double>(len);
CollectionsMarshal.SetCount(t, len);
CollectionsMarshal.SetCount(v, len);
var tSpan = CollectionsMarshal.AsSpan(t);
var vSpan = CollectionsMarshal.AsSpan(v);
Batch(actual.Values, predicted.Values, vSpan, period, delta);
actual.Times.CopyTo(tSpan);
return new TSeries(t, v);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static void Batch(ReadOnlySpan<double> actual, ReadOnlySpan<double> predicted, Span<double> output, int period, double delta = 1.345)
{
if (actual.Length != predicted.Length || actual.Length != output.Length)
throw new ArgumentException("All spans must have the same length", nameof(output));
if (period <= 0)
throw new ArgumentException("Period must be greater than 0", nameof(period));
if (delta <= 0)
throw new ArgumentException("Delta must be greater than 0", nameof(delta));
int len = actual.Length;
if (len == 0) return;
double halfDeltaSquared = 0.5 * delta * delta;
double negHalfDeltaSquared = -halfDeltaSquared;
const int StackAllocThreshold = 256;
Span<double> buffer = period <= StackAllocThreshold
? stackalloc double[period]
: new double[period];
// Pre-compute Huber losses using SIMD if available and data is clean
// Then apply rolling window average
Span<double> huberLosses = len <= StackAllocThreshold
? stackalloc double[len]
: new double[len];
ComputeHuberLosses(actual, predicted, huberLosses, delta, halfDeltaSquared, negHalfDeltaSquared);
// Apply rolling window average with O(1) per element
double sum = 0;
int bufferIndex = 0;
int warmupEnd = Math.Min(period, len);
for (int i = 0; i < warmupEnd; i++)
{
sum += huberLosses[i];
buffer[i] = huberLosses[i];
output[i] = sum / (i + 1);
}
int tickCount = 0;
for (int i = warmupEnd; i < len; i++)
{
double huberLoss = huberLosses[i];
sum = sum - buffer[bufferIndex] + huberLoss;
buffer[bufferIndex] = huberLoss;
bufferIndex++;
if (bufferIndex >= period) bufferIndex = 0;
output[i] = sum / period;
tickCount++;
if (tickCount >= ResyncInterval)
{
tickCount = 0;
double recalcSum = 0;
for (int k = 0; k < period; k++) recalcSum += buffer[k];
sum = recalcSum;
}
}
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static void ComputeHuberLosses(
ReadOnlySpan<double> actual,
ReadOnlySpan<double> predicted,
Span<double> huberLosses,
double delta,
double halfDeltaSquared,
double negHalfDeltaSquared)
{
int len = actual.Length;
double lastValidActual = 0;
double lastValidPredicted = 0;
// Find first valid values
for (int k = 0; k < len; k++)
{
if (double.IsFinite(actual[k])) { lastValidActual = actual[k]; break; }
}
for (int k = 0; k < len; k++)
{
if (double.IsFinite(predicted[k])) { lastValidPredicted = predicted[k]; break; }
}
// Try SIMD path for clean data (no NaN/Inf)
if (Avx2.IsSupported && len >= Vector256<double>.Count)
{
// Check if data is clean (no NaN/Inf) - sample check
bool dataClean = true;
int checkStep = Math.Max(1, len / 32);
for (int i = 0; i < len && dataClean; i += checkStep)
{
dataClean = double.IsFinite(actual[i]) && double.IsFinite(predicted[i]);
}
if (dataClean)
{
ComputeHuberLossesSimd(actual, predicted, huberLosses, delta, halfDeltaSquared, negHalfDeltaSquared);
return;
}
}
// Scalar fallback with NaN handling
ComputeHuberLossesScalar(actual, predicted, huberLosses, delta, negHalfDeltaSquared, lastValidActual, lastValidPredicted);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static void ComputeHuberLossesSimd(
ReadOnlySpan<double> actual,
ReadOnlySpan<double> predicted,
Span<double> huberLosses,
double delta,
double halfDeltaSquared,
double negHalfDeltaSquared)
{
int len = actual.Length;
int vectorSize = Vector256<double>.Count;
int vectorEnd = len - (len % vectorSize);
Vector256<double> deltaVec = Vector256.Create(delta);
Vector256<double> halfVec = Vector256.Create(0.5);
Vector256<double> negHalfDeltaSqVec = Vector256.Create(negHalfDeltaSquared);
int i = 0;
for (; i < vectorEnd; i += vectorSize)
{
Vector256<double> actVec = Vector256.LoadUnsafe(ref MemoryMarshal.GetReference(actual.Slice(i)));
Vector256<double> predVec = Vector256.LoadUnsafe(ref MemoryMarshal.GetReference(predicted.Slice(i)));
// error = actual - predicted
Vector256<double> errorVec = Avx.Subtract(actVec, predVec);
// absError = |error|
Vector256<double> absErrorVec = Avx.And(errorVec, Vector256.Create(~(1L << 63)).AsDouble());
// quadratic = 0.5 * error * error
Vector256<double> quadraticVec = Avx.Multiply(halfVec, Avx.Multiply(errorVec, errorVec));
// linear = delta * absError - halfDeltaSquared (using FMA)
Vector256<double> linearVec = Fma.IsSupported
? Fma.MultiplyAdd(deltaVec, absErrorVec, negHalfDeltaSqVec)
: Avx.Add(Avx.Multiply(deltaVec, absErrorVec), negHalfDeltaSqVec);
// mask = absError <= delta
Vector256<double> maskVec = Avx.CompareLessThanOrEqual(absErrorVec, deltaVec);
// result = mask ? quadratic : linear
Vector256<double> resultVec = Avx.BlendVariable(linearVec, quadraticVec, maskVec);
resultVec.StoreUnsafe(ref MemoryMarshal.GetReference(huberLosses.Slice(i)));
}
// Handle remainder with scalar
for (; i < len; i++)
{
double error = actual[i] - predicted[i];
double absError = Math.Abs(error);
huberLosses[i] = absError <= delta
? 0.5 * error * error
: Math.FusedMultiplyAdd(delta, absError, negHalfDeltaSquared);
}
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static void ComputeHuberLossesScalar(
ReadOnlySpan<double> actual,
ReadOnlySpan<double> predicted,
Span<double> huberLosses,
double delta,
double negHalfDeltaSquared,
double lastValidActual,
double lastValidPredicted)
{
int len = actual.Length;
for (int i = 0; i < len; i++)
{
double act = actual[i];
double pred = predicted[i];
if (double.IsFinite(act)) lastValidActual = act; else act = lastValidActual;
if (double.IsFinite(pred)) lastValidPredicted = pred; else pred = lastValidPredicted;
double error = act - pred;
double absError = Math.Abs(error);
huberLosses[i] = absError <= delta
? 0.5 * error * error
: Math.FusedMultiplyAdd(delta, absError, negHalfDeltaSquared);
}
}
}