/// <include file='FactorDocs.xml' path='factor_docs/message_op_class[@name="LogisticOp"]/message_doc[@name="LogisticAverageLogarithm(Gaussian)"]/*'/> public static Beta LogisticAverageLogarithm([Proper] Gaussian x) { double m, v; x.GetMeanAndVariance(out m, out v); #if true // for consistency with XAverageLogarithm double eLogOneMinusP = BernoulliFromLogOddsOp.AverageLogFactor(false, x); #else // E[log (1-sigma(x))] = E[log sigma(-x)] = -E[log(1+exp(x))] double eLogOneMinusP = -MMath.Log1PlusExpGaussian(m, v); #endif // E[log sigma(x)] = -E[log(1+exp(-x))] = -E[log(1+exp(x))-x] = -E[log(1+exp(x))] + E[x] double eLogP = eLogOneMinusP + m; return(Beta.FromMeanLogs(eLogP, eLogOneMinusP)); }
/// <include file='FactorDocs.xml' path='factor_docs/message_op_class[@name="LogisticOp"]/message_doc[@name="XAverageLogarithm(Beta, Gaussian, Gaussian)"]/*'/> public static Gaussian XAverageLogarithm([SkipIfUniform] Beta logistic, [Proper, SkipIfUniform] Gaussian x, Gaussian to_X) { if (logistic.IsPointMass) { return(XAverageLogarithm(logistic.Point)); } // f(x) = sigma(x)^(a-1) sigma(-x)^(b-1) // = sigma(x)^(a+b-2) exp(-x(b-1)) // since sigma(-x) = sigma(x) exp(-x) double a = logistic.TrueCount; double b = logistic.FalseCount; double scale = a + b - 2; if (scale == 0.0) { return(Gaussian.Uniform()); } double shift = -(b - 1); Gaussian toLogOddsPrev = Gaussian.FromNatural((to_X.MeanTimesPrecision - shift) / scale, to_X.Precision / scale); Gaussian toLogOdds = BernoulliFromLogOddsOp.LogOddsAverageLogarithm(true, x, toLogOddsPrev); return(Gaussian.FromNatural(scale * toLogOdds.MeanTimesPrecision + shift, scale * toLogOdds.Precision)); }
/// <include file='FactorDocs.xml' path='factor_docs/message_op_class[@name="LogisticOp"]/message_doc[@name="FalseMsg(Beta, Gaussian, Gaussian)"]/*'/> public static Gaussian FalseMsg([SkipIfUniform] Beta logistic, [Proper] Gaussian x, Gaussian falseMsg) { // falseMsg approximates sigma(-x) // logistic(sigma(x)) N(x;m,v) // = sigma(x)^(a-1) sigma(-x)^(b-1) N(x;m,v) // = e^((a-1)x) sigma(-x)^(a+b-2) N(x;m,v) // = sigma(-x)^(a+b-2) N(x;m+(a-1)v,v) exp((a-1)m + (a-1)^2 v/2) // = sigma(-x) (prior) // where prior = sigma(-x)^(a+b-3) N(x;m+(a-1)v,v) double tc1 = logistic.TrueCount - 1; double fc1 = logistic.FalseCount - 1; double m, v; x.GetMeanAndVariance(out m, out v); if (tc1 + fc1 == 0) { falseMsg.SetToUniform(); return(falseMsg); } else if (tc1 + fc1 < 0) { // power EP update, using 1/sigma(-x) as the factor Gaussian prior = new Gaussian(m + tc1 * v, v) * (falseMsg ^ (tc1 + fc1 + 1)); double mprior, vprior; prior.GetMeanAndVariance(out mprior, out vprior); // posterior moments can be computed exactly double w = MMath.Logistic(mprior + 0.5 * vprior); Gaussian post = new Gaussian(mprior + w * vprior, vprior * (1 + w * (1 - w) * vprior)); return(prior / post); } else { // power EP update Gaussian prior = new Gaussian(m + tc1 * v, v) * (falseMsg ^ (tc1 + fc1 - 1)); Gaussian newMsg = BernoulliFromLogOddsOp.LogOddsAverageConditional(false, prior); //Console.WriteLine("prior = {0}, falseMsg = {1}, newMsg = {2}", prior, falseMsg, newMsg); if (string.Empty.Length == 0) { // adaptive damping scheme Gaussian ratio = newMsg / falseMsg; if ((ratio.MeanTimesPrecision < 0 && prior.MeanTimesPrecision > 0) || (ratio.MeanTimesPrecision > 0 && prior.MeanTimesPrecision < 0)) { // if the update would change the sign of the mean, take a fractional step so that the new prior has exactly zero mean // newMsg = falseMsg * (ratio^step) // newPrior = prior * (ratio^step)^(tc1+fc1-1) // 0 = prior.mp + ratio.mp*step*(tc1+fc1-1) double step = -prior.MeanTimesPrecision / (ratio.MeanTimesPrecision * (tc1 + fc1 - 1)); if (step > 0 && step < 1) { newMsg = falseMsg * (ratio ^ step); // check that newPrior has zero mean //Gaussian newPrior = prior * ((ratio^step)^(tc1+fc1-1)); //Console.WriteLine(newPrior); } } } else { for (int iter = 0; iter < 10; iter++) { newMsg = falseMsg * ((newMsg / falseMsg) ^ 0.5); falseMsg = newMsg; //Console.WriteLine("prior = {0}, falseMsg = {1}, newMsg = {2}", prior, falseMsg, newMsg); prior = new Gaussian(m + tc1 * v, v) * (falseMsg ^ (tc1 + fc1 - 1)); newMsg = BernoulliFromLogOddsOp.LogOddsAverageConditional(false, prior); } } return(newMsg); } }