Exemplo n.º 1
0
/*Beta function
Special cases:
z<0: NaN
w<0: NaN
*/
func Beta(z float64, w float64) float64 {
	if z < 0 || w < 0 {
		return math.NaN()
	}
	a, _ := math.Lgamma(z)
	b, _ := math.Lgamma(w)
	c, _ := math.Lgamma(z + w)
	return math.Exp(a + b - c)
}
Exemplo n.º 2
0
func lfastchoose2(n, k float64) (float64, int) {
	// mathematically the same as lfastchoose()
	// less stable typically, but useful if n-k+1 < 0
	//	r := lgammafn_sign(n-k+1, s_choose)
	r, s_choose := math.Lgamma(n - k + 1)
	p, _ := math.Lgamma(n + 1)
	q, _ := math.Lgamma(k + 1)
	return p - q - r, s_choose
}
Exemplo n.º 3
0
func (sampler *Sampler) updateComponentE(targetEvent int) (lgamma float64) {
	var tPositive, tNegative, tNormalize float64
	for k := 0; k < numTop; k++ {
		tPositive, _ = math.Lgamma(float64(sampler.Model.Eventtype_histogram[k]) + sampler.eventPrior)
		tNegative, _ = math.Lgamma(float64((sampler.Model.NumESDs)-sampler.Model.Eventtype_histogram[k]) + sampler.eventPrior)
		tNormalize, _ = math.Lgamma(float64(sampler.Model.NumESDs) + 2*sampler.eventPrior)
		lgamma += ((tPositive + tNegative) - tNormalize)
	}
	return lgamma
}
Exemplo n.º 4
0
func (sampler *Sampler) updateComponentP(participantID, eventID int) (lgamma float64) {
	// for each alternative participanttype
	var pPositive, pNegative, pNormalize float64
	for i := 0; i < numPar; i++ {
		pPositive, _ = math.Lgamma(float64(sampler.Model.Participanttype_eventtype_histogram[i][eventID]) + sampler.participantPrior)
		pNegative, _ = math.Lgamma(float64(sampler.Model.Eventtype_histogram[eventID]-sampler.Model.Participanttype_eventtype_histogram[i][eventID]) + sampler.participantPrior)
		pNormalize, _ = math.Lgamma(float64(sampler.Model.Eventtype_histogram[eventID]) + 2*sampler.participantPrior)
		lgamma += ((pPositive + pNegative) - pNormalize)
	}
	return
}
Exemplo n.º 5
0
// Calculates Ix(a, b). Borrowed from
// NUMERICAL RECIPES IN FORTRAN 77: THE ART OF SCIENTIFIC COMPUTING
// ISBN (0-521-43064-Z)
func betaInc(a, b, x float64) float64 {
	lgab, _ := math.Lgamma(a + b)
	lga, _ := math.Lgamma(a)
	lgb, _ := math.Lgamma(b)
	exp := lgab - lga - lgb + a*math.Log(x) + b*math.Log(1.0-x)
	bt := math.Exp(exp)
	if x < (a+1.0)/(a+b+2.0) {
		return bt * betaCF(a, b, x) / a
	} else {
		return 1.0 - bt*betaCF(b, a, 1.0-x)/b
	}
}
Exemplo n.º 6
0
// LogGeneralizedBinomial returns the log of the generalized binomial coefficient.
// See GeneralizedBinomial for more information.
func LogGeneralizedBinomial(n, k float64) float64 {
	if n < 0 || k < 0 {
		panic(badNegInput)
	}
	if n < k {
		panic(badSetSize)
	}
	a, _ := math.Lgamma(n + 1)
	b, _ := math.Lgamma(k + 1)
	c, _ := math.Lgamma(n - k + 1)
	return a - b - c
}
Exemplo n.º 7
0
func main() {
	for true {
		r := bufio.NewReader(os.Stdin)
		s, err := r.ReadString('\n')
		if err == os.EOF {
			break
		}
		s = strings.TrimRight(s, "\n")
		a := strings.Split(s, " ")
		f := a[0]
		x, err := strconv.Atof64(a[1])
		switch f {
		case "erf":
			fmt.Println(math.Erf(x))
		case "expm1":
			fmt.Println(math.Expm1(x))
		case "phi":
			fmt.Println(phi.Phi(x))
		case "NormalCDFInverse":
			fmt.Println(normal_cdf_inverse.NormalCDFInverse(x))
		case "Gamma":
			fmt.Println(math.Gamma(x))
		case "LogGamma":
			r, _ := math.Lgamma(x)
			fmt.Println(r)
		case "LogFactorial":
			fmt.Println(log_factorial.LogFactorial(int(x)))
		default:
			fmt.Println("Unknown function: " + f)
			return
		}
	}
}
Exemplo n.º 8
0
// Lower incomplete gamma.
func lgamma(x, s float64, regularized bool) float64 {
	if x == 0 {
		return 0
	}
	if x < 0 || s <= 0 {
		return math.NaN()
	}

	if x > 1.1 && x > s {
		if regularized {
			return 1.0 - ugamma(x, s, regularized)
		}
		return math.Gamma(s) - ugamma(x, s, regularized)
	}

	var ft float64
	r := s
	c := 1.0
	pws := 1.0

	if regularized {
		logg, _ := math.Lgamma(s)
		ft = s*math.Log(x) - x - logg
	} else {
		ft = s*math.Log(x) - x
	}
	ft = math.Exp(ft)
	for c/pws > eps {
		r++
		c *= x / r
		pws += c
	}
	return pws * ft / s
}
Exemplo n.º 9
0
// Upper incomplete gamma.
func ugamma(x, s float64, regularized bool) float64 {
	if x <= 1.1 || x <= s {
		if regularized {
			return 1 - lgamma(x, s, regularized)
		}
		return math.Gamma(s) - lgamma(x, s, regularized)
	}

	f := 1.0 + x - s
	C := f
	D := 0.0
	var a, b, chg float64

	for i := 1; i < 10000; i++ {
		a = float64(i) * (s - float64(i))
		b = float64(i<<1) + 1.0 + x - s
		D = b + a*D
		C = b + a/C
		D = 1.0 / D
		chg = C * D
		f *= chg
		if math.Abs(chg-1) < eps {
			break
		}
	}
	if regularized {
		logg, _ := math.Lgamma(s)
		return math.Exp(s*math.Log(x) - x - logg - math.Log(f))
	}
	return math.Exp(s*math.Log(x) - x - math.Log(f))
}
Exemplo n.º 10
0
// Skewness returns the skewness of the distribution.
func (w Weibull) Skewness() float64 {
	stdDev := w.StdDev()
	firstGamma, firstGammaSign := math.Lgamma(1 + 3/w.K)
	logFirst := firstGamma + 3*(math.Log(w.Lambda)-math.Log(stdDev))
	logSecond := math.Log(3) + math.Log(w.Mean()) + 2*math.Log(stdDev) - 3*math.Log(stdDev)
	logThird := 3 * (math.Log(w.Mean()) - math.Log(stdDev))
	return float64(firstGammaSign)*math.Exp(logFirst) - math.Exp(logSecond) - math.Exp(logThird)
}
Exemplo n.º 11
0
/*Normalized Gamma Function
(or Complementary Incomplete Gamma Function)

Equal to Gamma(a, x)/Gamma(a)
Evaluated by Legendre's continued fraction

Special Cases:
G(0, 0) = Infinity
G(0, positive) = 1.0
G(0, negative) = 1/(-a)
*/
func G(a float64, x float64) float64 {
	if x == 0 {
		if a == 0 {
			return math.Inf(1)
		} else if a > 0 {
			return 1.0
		} else if a < 0 {
			return 1.0 / math.Abs(a)
		}
	}

	//Evaluate Legendre's continued fraction
	//using Lentz's algorithm
	//Shift from Thomson and Barnett
	//Continued fraction:
	//http://functions.wolfram.com/GammaBetaErf/GammaRegularized/10/0003/

	b0 := x + 1.0 - a
	C := 1.0 / (10E-30)
	D := 1.0 / b0
	if b0 == 0 {
		D = 10E30
	}
	f := D
	//numerator
	an := func(n int) float64 {
		if n == 0 {
			return 1.0
		}
		return -1.0 * float64(n) * (float64(n) - a)
	}
	//denominator
	bn := func(n int) float64 {
		return x + (float64(2*n + 1)) - a
	}
	//Lentz's algorithm until machine precision or 1,000 iterations
	for j := 1; j < 1000; j++ {
		D = bn(j) + an(j)*D
		if math.Abs(D) < 10E-20 {
			D = 10E-30
		}
		C = bn(j) + an(j)/C
		if math.Abs(C) < 10E-20 {
			C = 10E-30
		}
		D = 1.0 / D
		del := D * C
		f *= del
		if math.Abs(del-1.0) < 10E-15 {
			break
		}
	}
	lnGa, _ := math.Lgamma(a)
	return f * math.Exp(-x+a*math.Log(x)-lnGa)
}
Exemplo n.º 12
0
// compute document likelihood of the events in the current esd
// all participant labelings will stay constant -> no need to compute them!
func (sampler *Sampler) documentLikelihood(label Label) float64 {
	var wordTypeFactor, wordFactor, wordNorm float64
	var typeWordTotal int
	documentLikelihood := 0.0
	// iterate over eventtypes
	for k := 0; k < numTop; k++ {
		wordFactor = 0.0
		typeWordTotal = 0
		// iterate over terms in event-vocab
		for term, histogram := range sampler.Model.Word_eventtype_histogram {
			typeWordTotal += histogram[k]
			// compute LGamma(N(word,event) + prior + udpate)
			wordTypeFactor, _ = math.Lgamma(float64(histogram[k]) + sampler.EventlmPriors[k][term])
			wordFactor += wordTypeFactor
		}
		// normalize LGamma(N(words_by_event) + V*prior + total_update)
		wordNorm, _ = math.Lgamma(float64(typeWordTotal) + sum(sampler.EventlmPriors[k]))
		documentLikelihood += (wordFactor - wordNorm)
	}
	return documentLikelihood
}
Exemplo n.º 13
0
// compute document likelihood of the participant realization in question, given the proposed label
// all event doc likelihoods will stay constant w.r.t. change -> no need to compute them!
func (sampler *Sampler) documentLikelihoodP(event int, participant int, label Label) float64 {
	var wordTypeFactor, wordFactor, wordNorm float64
	var typeWordTotal /*, update*/ int
	documentLikelihood := 0.0
	// iterate over participanttypes
	for i := 0; i < numPar; i++ {
		wordFactor = 0.0
		typeWordTotal = 0
		// iterate over terms in participant vocab
		for term, histogram := range sampler.Model.Word_participanttype_histogram {
			typeWordTotal += histogram[i]
			// set 'update' according to the number of times term is present in current particip descr
			// compute LGamma(N(word,part) + prior + update)
			wordTypeFactor, _ = math.Lgamma(float64(histogram[i]) + sampler.ParticipantlmPriors[i][term])
			wordFactor += wordTypeFactor
		}
		// normalize
		wordNorm, _ = math.Lgamma(float64(typeWordTotal) + sum(sampler.ParticipantlmPriors[i]))
		documentLikelihood += (wordFactor - wordNorm)
	}
	return documentLikelihood
}
Exemplo n.º 14
0
//Gamma Distribution PDF
func (g *Gamma) PDF(x float64) float64 {
	if x <= 0 {
		return math.NaN()
	}
	if math.IsInf(x, 1) {
		return 1.0
	} else if math.IsInf(x, -1) {
		return 0.0
	}
	lga, _ := math.Lgamma(g.alpha)
	logp := (g.alpha * math.Log(g.beta)) + ((g.alpha - 1.0) * math.Log(x)) - (x * g.beta) - lga
	return math.Exp(logp)
}
Exemplo n.º 15
0
func TestLnGamma(t *testing.T) {
	acc := 0.0000001
	check := func(x, y float64) bool {
		if false {
			return x == y
		}
		return math.Abs(x-y) < acc
	}
	for i := 0; i < 100; i++ {
		x := NextGamma(10, 10)
		g1 := LnΓ(x)
		g2, _ := math.Lgamma(x)
		if !check(g1, g2) {
			t.Error(fmt.Sprintf("For %v: %v vs %v", x, g1, g2))
		}
	}
	//var start int64
	Seed(10)
	start := time.Now()
	for i := 0; i < 1e6; i++ {
		x := NextGamma(10, 10)
		math.Lgamma(x)
	}
	now := time.Now()
	duration2 := float64(now.Sub(start)) / 1e9

	//duration2 := float64(time.Now()-start) / 1e9
	Seed(10)
	start = time.Now()
	for i := 0; i < 1e6; i++ {
		x := NextGamma(10, 10)
		LnΓ(x)
	}
	now = time.Now()
	duration1 := float64(now.Sub(start)) / 1e9
	fmt.Printf("Mine was %f\nTheirs was %f\n", duration1, duration2)
}
Exemplo n.º 16
0
// incGammaSeries computes the incomplete gamma function via the
// series sum:
//
// gamma(a, x) = exp(-x) x^a sum_n (Gamma(a) / Gamma(a + 1 + n)) x^n.
func incGammaSeries(a, x float64) float64 {
	an := a
	termVal := 1 / an
	sum := 1 / an

	for n := 0; n < ConvergenceIters; n++ {
		an++
		termVal *= x / float64(an)
		sum += termVal
		if CloseEnough(sum, termVal) {
			// Note that x^a = exp(a * log(x)).
			lg, _ := math.Lgamma(a)
			return sum * math.Exp(-x+a*math.Log(x)-lg)
		}
	}

	fmtStr := "incGammaSeries(%g, %g) failed to converge"
	panic(fmt.Sprintf(fmtStr, a, x))
}
Exemplo n.º 17
0
// incGammaContinuedFraction computes the incomplete gamma function
// via the contiued fraction:
//
// exp(-x) x^ a (1 / (x + 1 - a - (1 (1 - a)) / (x + 3 - a - ...))).
//
// This is implemented via modified Lentz's method.
func incGammaContinuedFraction(a, x float64) float64 {
	minValue := math.SmallestNonzeroFloat64

	// Start computing terms at n = 1.

	// c_n = A_n / A_n-1. Note that A_0 = a_0 = 0.
	c := math.MaxFloat64

	// c_n = B_n-1 / B_n. Note that B_0 = b_0 = 1.
	b := x + 1 - a
	d := 1 / b

	fracEst := d

	for n := 1; n <= ConvergenceIters; n++ {
		an := -float64(n) * (float64(n) - a)
		b += 2.0

		d = an*d + b
		if math.Abs(d) < minValue {
			d = minValue
		}

		c = an/c + b
		if math.Abs(c) < minValue {
			c = minValue
		}

		d = 1 / d
		diff := d * c
		fracEst *= diff

		if CloseEnough(1, diff-1) {
			lg, _ := math.Lgamma(a)
			return fracEst * math.Exp(-x+a*math.Log(x)-lg)
		}
	}

	fmtStr := "incGammaContinuedFraction(%g, %g) failed to converge"
	panic(fmt.Sprintf(fmtStr, a, x))
}
Exemplo n.º 18
0
//Based on similar code in rf-ace (Apache 2.0, Timo Erkkilä)
func lgamma(x float64) float64 {
	v, _ := math.Lgamma(x)
	//v := math.Log(math.Abs(math.Gamma(x)))
	return v
}
Exemplo n.º 19
0
func lchoose(n, k int) float64 {
	a, _ := math.Lgamma(float64(n + 1))
	b, _ := math.Lgamma(float64(k + 1))
	c, _ := math.Lgamma(float64(n - k + 1))
	return a - b - c
}
Exemplo n.º 20
0
// Lgamma computes the log gamma function, but ignores domain errors.
func Lgamma(x float64) (y float64) {
	y, _ = math.Lgamma(x)
	return
}
Exemplo n.º 21
0
func Lgamma(x float64) float64 {
	val, _ := math.Lgamma(x)
	return val
}
Exemplo n.º 22
0
func (l *LdaModeler) model_one_pass(tokenss_ptr *[]util.Document, seed int, initial_topicss [][]int) *ModelResults {
	// log.Printf("we're in %p\n", l)
	// log.Printf("from %p: address of initial_topicss: %p\n", l, &initial_topicss)
	tokenss := *tokenss_ptr
	ntopics := l.ntopics
	ndocuments := len(tokenss)
	range_ntopics := util.PyRange(ntopics)
	range_ndocuments := util.PyRange(ndocuments)

	// set up a random number generator
	randomizer := rand.New(rand.NewSource(int64(len(tokenss) + seed)))

	// topicss is an document-token matrix where the values are topic assignments.
	topicss := make([][]int, len(tokenss))

	if initial_topicss == nil {
		// set up new, randomly initialized matrix
		for i, doc := range tokenss {
			doc_topics := make([]int, len(doc))
			for j, _ := range doc { // randomly assign each token in doc to a random topic
				doc_topics[j] = randomizer.Intn(ntopics)
			}
			topicss[i] = doc_topics
		}
	} else {
		// we were passed in a topic matrix; let's just set up our own local copy
		for i, doc := range initial_topicss {
			// NOTE: We *cannot* just say topicss[i] = doc- shallow copy vs. deep copy, pointers, etc. etc.- here there be possibilities for threading havoc!
			topicss[i] = make([]int, len(doc))
			copy(topicss[i], doc)
		}
	}

	// compute doc-topic representation counts and topic-word representation counts
	document_tokens_counts := make([]int, len(tokenss))
	for idx, doc := range tokenss {
		document_tokens_counts[idx] = len(doc)
	}

	document_topics_counts := make([][]int, len(tokenss))
	topic_words_counts := make([]map[util.Token]int, ntopics) // maps in Go will, if asked for an element that doesn't exist, give the zero-value for that type (as well as an optional second return value indicating whether it was found or not) http://golang.org/doc/effective_go.html#maps
	total_topic_counts := make([]int, ntopics)                // n.b.: the values in a "fresh" just-made slice are the zero-value for that type.

	for i, tokens := range tokenss { // for each document
		topics := topicss[i]
		counts := make([]int, ntopics)
		for j, token := range tokens { // for each token
			topic := topics[j]
			counts[topic] += 1 // count of topic mentions in this document
			if topic_words_counts[topic] == nil {
				topic_words_counts[topic] = map[util.Token]int{}
			}
			topic_words_counts[topic][token] += 1 // count tokens mentions for this topic
			total_topic_counts[topic] += 1        // total count of this topic
		}
		document_topics_counts[i] = counts // = append(document_topics_counts, counts)
	}

	all_keys := make([][]util.Token, len(range_ntopics)) // a set of per-topic vocab lists
	for i := range range_ntopics {
		all_keys[i] = util.KeysFromMap(topic_words_counts[i])
	}
	all_tokens := util.SetFromLists(all_keys)

	// Dirichlet smoothing parameters
	alpha := l.alpha
	beta := l.beta
	W := len(all_tokens) // size of vocab: num of possible unique words in each topic
	T := ntopics         // num topics
	betaW := beta * float64(W)
	alphaT := alpha * float64(T)
	// maxT := ntopics - 1
	// uniform_random_func := randomizer.Float64 // seems to be equivalent to Python's random()- uniform dist between [0.0, 1.0]

	// loop over all docs and all tokens, resampling topic assignments & adjusting counts
	proportional_probabilities := make([]float64, ntopics) // probability of each topic

	fixups := make([]int, ntopics) // which topics need count adjustment for current token
	for iteration := 0; iteration < l.iterations; iteration++ {
		change_count := 0
		for t_idx, tokens := range tokenss { // for each document
			topics := topicss[t_idx]
			document_index := range_ndocuments[t_idx]
			current_document_topics_counts := document_topics_counts[document_index]
			current_document_tokens_count := document_tokens_counts[document_index]
			n_di_minus_i := float64(current_document_tokens_count - 1)
			for token_index, token := range tokens { // for each token
				// Based on:
				// Griffiths TL, Steyvers M. Finding scientific topics.
				// Proceedings of the National Academy of Sciences of the United States of America. 2004;101(Suppl 1):5228-5235.

				// get topic assignment for current token:
				topic := topics[token_index]

				// compute conditional probabilities for each topic,
				// the "fixups" list is an optimization to avoid branching.
				fixups[topic] = 1

				total_proportional_probabilities := 0.0
				for _, j := range range_ntopics { // for each topic
					fixup := fixups[j]

					n_wi_minus_i_j := float64(topic_words_counts[j][token] - fixup) // most of the time, fixup will be zero

					n_di_minus_i_j := float64(current_document_topics_counts[j] - fixup) // ditto
					n_dot_minus_i_j := float64(total_topic_counts[j] - fixup)

					// eq. 5 from above paper
					p_token_topic := (n_wi_minus_i_j + beta) / (n_dot_minus_i_j + betaW)
					p_topic_document := (n_di_minus_i_j + alpha) / (n_di_minus_i + alphaT)
					p := p_topic_document * p_token_topic
					proportional_probabilities[j] = p
					total_proportional_probabilities += p
				} // end for topics
				fixups[topic] = 0

				// resample current token topic, integrate the inline version of resample function
				new_topic := l.resample(randomizer.Float64(), proportional_probabilities, total_proportional_probabilities)

				// update assignments & counts
				if new_topic != topic {
					// update topic label for this token:
					topics[token_index] = new_topic

					// update total topic counts:
					total_topic_counts[topic] -= 1
					total_topic_counts[new_topic] += 1

					// update document-topic counts
					current_document_topics_counts[topic] -= 1
					current_document_topics_counts[new_topic] += 1

					topic_words_counts[topic][token] -= 1
					topic_words_counts[new_topic][token] += 1

					// count changes for this pass
					change_count += 1
				}
			} // end for tokens
		} // end for document
		// log.Printf("LDA - iteration %d resulted in %d changes.\n", iteration, change_count)
		if iteration%100 == 0 {
			log.Printf("LDA - iteration %d of %d.\n", iteration, l.iterations)
		}
	} // for iterations

	// document-topic assignments (theta_hat_d_j)
	theta_hat_ds := make([][]float64, ndocuments)

	for document_index := 0; document_index < ndocuments; document_index++ {
		document_token_count := document_tokens_counts[document_index]
		theta_hat_d := make([]float64, ntopics)
		document_topics_count := document_topics_counts[document_index]
		if document_token_count > 0 {
			for j := 0; j < ntopics; j++ {
				p := (float64(document_topics_count[j]) + alpha) / (float64(document_token_count) + alphaT)
				theta_hat_d[j] = p
			}
		} else {
			// degenerate document with no tokens- equal prob of all topics
			// temp := float64((1.0 / ntopics) * ntopics)
			for j := range range_ntopics {
				theta_hat_d[j] = float64(1.0 / ntopics) //temp
			}
		}
		theta_hat_ds[document_index] = theta_hat_d
	}

	// compute topic-token assignments (phi_hat_w_j in paper)
	phi_hats := make([]util.TokenProbMap, ntopics)
	for t := 0; t < ntopics; t++ { // for each topic
		dx := util.TokenProbMap{}
		for token, top_tok_count := range topic_words_counts[t] { // for each token
			dx[token] = util.Probability((float64(top_tok_count) + beta) / (float64(total_topic_counts[t]) + betaW))
		}
		phi_hats[t] = dx
	}

	// compute log-likelihood of tokens given topic model; Eq. 2 in Steyvers paper
	part_1, _ := math.Lgamma(float64(W) * beta) // note that Lgamma returns both the gamma, as well as a sign indicator- we don't care about the latter here
	part_2, _ := math.Lgamma(beta)
	log_likelihood := float64(T) * (part_1 - (float64(W) * part_2))
	for t := 0; t < ntopics; t++ {
		for _, w := range all_tokens {
			n_t_w := topic_words_counts[t][w]
			ntw_gamma, _ := math.Lgamma(float64(n_t_w) + beta)
			log_likelihood += ntw_gamma
		}
		n_dot_t := total_topic_counts[t]
		ndt_gamma, _ := math.Lgamma(float64(n_dot_t) + betaW)
		log_likelihood -= ndt_gamma
	}
	log.Printf("LDA - log-likelihood of data given model: %0.8e\n", log_likelihood)

	// sum over samples:
	// See definition in:
	// Chemudugunta C, Steyvers PSM. Modeling General and Specific Aspects of Documents with a Probabilistic Topic Model.
	// In: Advances in Neural Information Processing Systems 19: Proceedings of the 2006 Conference. MIT Press; 2007.  p. 241.
	//
	// modified implementation to add the logs of the P(tokens)'s rather than multiply
	// the P(tokens) and then take the log in order to avoid underflowing to zero.
	perplexity := 0.0
	ntokens := 0

	for doc_idx, tokens := range tokenss { // each document
		theta_hat_d := theta_hat_ds[doc_idx]
		for _, w := range tokens { // each token
			temp_phi_hat_theta_hat := make([]float64, len(range_ntopics))
			for j, z := range range_ntopics { // each topic
				temp_phi_hat_theta_hat[j] = float64(phi_hats[z][w]) * theta_hat_d[z]
			}
			perplexity += math.Log2(util.SumFloat(temp_phi_hat_theta_hat))
			ntokens += 1
		}
	}

	perplexity = math.Pow(2.0, (-perplexity / float64(ntokens)))
	log.Printf("LDA - mean sample token perplexity of data given model: %0.4f\n", perplexity)

	// save final snapshot of token-topic assignment snapshot
	if l.last_model_token_topic_sample_assignments == nil {
		l.last_model_token_topic_sample_assignments = topicss
	}

	// return results sample
	return &ModelResults{&theta_hat_ds, &phi_hats, log_likelihood, perplexity}

}
Exemplo n.º 23
0
func Lgamma(f float32) (lgamma float32, sign int) {
	lgamma64, sign := math.Lgamma(float64(f))
	lgamma = float32(lgamma64)
	return
}
Exemplo n.º 24
0
func lgamma(x float64) float64 {
	y, _ := math.Lgamma(x)
	return y
}