Exemplo n.º 1
0
 /// <summary>
 /// Returns a value indicating whether the specified number evaluates to positive infinity
 /// </summary>
 /// <param name="number">a floating point number</param>
 /// <returns>a boolean</returns>
 public static bool IsPositiveInfinity(Real number)
 {
     return(Numeric.IsPositiveInfinity((Numeric)number));
 }
Exemplo n.º 2
0
            /// <summary>
            /// An implementation of the line search for the Wolfe conditions, from Nocedal &amp; Wright
            /// </summary>
            internal virtual bool LineSearch(IChannel ch, bool force)
            {
                Contracts.AssertValue(ch);
                Float dirDeriv = VectorUtils.DotProduct(ref _dir, ref _grad);

                if (dirDeriv == 0)
                {
                    throw ch.Process(new PrematureConvergenceException(this, "Directional derivative is zero. You may be sitting on the optimum."));
                }

                // if a non-descent direction is chosen, the line search will break anyway, so throw here
                // The most likely reasons for this is a bug in your function's gradient computation,
                ch.Check(dirDeriv < 0, "L-BFGS chose a non-descent direction.");

                Float c1 = (Float)1e-4 * dirDeriv;
                Float c2 = (Float)0.9 * dirDeriv;

                Float alpha = (Iter == 1 ? (1 / VectorUtils.Norm(_dir)) : 1);

                PointValueDeriv last = new PointValueDeriv(0, LastValue, dirDeriv);
                PointValueDeriv aLo  = new PointValueDeriv();
                PointValueDeriv aHi  = new PointValueDeriv();

                // initial bracketing phase
                while (true)
                {
                    VectorUtils.AddMultInto(ref _x, alpha, ref _dir, ref _newX);
                    if (EnforceNonNegativity)
                    {
                        VBufferUtils.Apply(ref _newX, delegate(int ind, ref Float newXval)
                        {
                            if (newXval < 0.0)
                            {
                                newXval = 0;
                            }
                        });
                    }

                    Value = Eval(ref _newX, ref _newGrad);
                    GradientCalculations++;
                    if (Float.IsPositiveInfinity(Value))
                    {
                        alpha /= 2;
                        continue;
                    }

                    if (!FloatUtils.IsFinite(Value))
                    {
                        throw ch.Except("Optimizer unable to proceed with loss function yielding {0}", Value);
                    }

                    dirDeriv = VectorUtils.DotProduct(ref _dir, ref _newGrad);
                    PointValueDeriv curr = new PointValueDeriv(alpha, Value, dirDeriv);

                    if ((curr.V > LastValue + c1 * alpha) || (last.A > 0 && curr.V >= last.V))
                    {
                        aLo = last;
                        aHi = curr;
                        break;
                    }
                    else if (Math.Abs(curr.D) <= -c2)
                    {
                        return(true);
                    }
                    else if (curr.D >= 0)
                    {
                        aLo = curr;
                        aHi = last;
                        break;
                    }

                    last = curr;
                    if (alpha == 0)
                    {
                        alpha = Float.Epsilon; // Robust to divisional underflow.
                    }
                    else
                    {
                        alpha *= 2;
                    }
                }

                Float minChange = (Float)0.01;
                int   maxSteps  = 10;

                // this loop is the "zoom" procedure described in Nocedal & Wright
                for (int step = 0; ; ++step)
                {
                    if (step == maxSteps && !force)
                    {
                        return(false);
                    }

                    PointValueDeriv left  = aLo.A < aHi.A ? aLo : aHi;
                    PointValueDeriv right = aLo.A < aHi.A ? aHi : aLo;
                    if (left.D > 0 && right.D < 0)
                    {
                        // interpolating cubic would have max in range, not min (can this happen?)
                        // set a to the one with smaller value
                        alpha = aLo.V < aHi.V ? aLo.A : aHi.A;
                    }
                    else
                    {
                        alpha = CubicInterp(aLo, aHi);
                        if (Float.IsNaN(alpha) || Float.IsInfinity(alpha))
                        {
                            alpha = (aLo.A + aHi.A) / 2;
                        }
                    }

                    // this is to ensure that the new point is within bounds
                    // and that the change is reasonably sized
                    Float ub = (minChange * left.A + (1 - minChange) * right.A);
                    if (alpha > ub)
                    {
                        alpha = ub;
                    }
                    Float lb = (minChange * right.A + (1 - minChange) * left.A);
                    if (alpha < lb)
                    {
                        alpha = lb;
                    }

                    VectorUtils.AddMultInto(ref _x, alpha, ref _dir, ref _newX);
                    if (EnforceNonNegativity)
                    {
                        VBufferUtils.Apply(ref _newX, delegate(int ind, ref Float newXval)
                        {
                            if (newXval < 0.0)
                            {
                                newXval = 0;
                            }
                        });
                    }

                    Value = Eval(ref _newX, ref _newGrad);
                    GradientCalculations++;
                    if (!FloatUtils.IsFinite(Value))
                    {
                        throw ch.Except("Optimizer unable to proceed with loss function yielding {0}", Value);
                    }
                    dirDeriv = VectorUtils.DotProduct(ref _dir, ref _newGrad);

                    PointValueDeriv curr = new PointValueDeriv(alpha, Value, dirDeriv);

                    if ((curr.V > LastValue + c1 * alpha) || (curr.V >= aLo.V))
                    {
                        if (aHi.A == curr.A)
                        {
                            if (force)
                            {
                                throw ch.Process(new PrematureConvergenceException(this, "Step size interval numerically zero."));
                            }
                            else
                            {
                                return(false);
                            }
                        }
                        aHi = curr;
                    }
                    else if (Math.Abs(curr.D) <= -c2)
                    {
                        return(true);
                    }
                    else
                    {
                        if (curr.D * (aHi.A - aLo.A) >= 0)
                        {
                            aHi = aLo;
                        }
                        if (aLo.A == curr.A)
                        {
                            if (force)
                            {
                                throw ch.Process(new PrematureConvergenceException(this, "Step size interval numerically zero."));
                            }
                            else
                            {
                                return(false);
                            }
                        }
                        aLo = curr;
                    }
                }
            }