func Train(data []Datum) *NaiveBayes {
class := counter.New(0.0)
features := make(map[string]*counter.Counter)
for _, datum := range data {
class.Incr(datum.class)
for _, f := range datum.features {
dist, ok := features[f]
if !ok {
dist = counter.New(0.0)
features[f] = dist
}
dist.Incr(datum.class)
}
}
class.LogNormalize()
for _, dist := range features {
dist.LogNormalize()
}
frozenFeatures := frozencounter.FreezeMap(features)
var keyset *frozencounter.KeySet
for _, dist := range frozenFeatures {
keyset = dist.Keys
}
frozenClass := frozencounter.FreezeWithKeySet(class, keyset)
return &NaiveBayes{FeatureLogDistributions: frozenFeatures, ClassLogPrior: frozenClass}
}
func tally(data []Datum) (counts *frozencounter.CounterVector, features *frozencounter.KeySet, labels []string) {
rawCounts := map[string]*counter.Counter{}
datumCounts := []*counter.Counter{}
for _, datum := range data {
if rawCounts[datum.class] == nil {
rawCounts[datum.class] = counter.New(0.0)
}
c := counter.New(0.0)
for _, f := range datum.features {
rawCounts[datum.class].Incr(f)
c.Incr(f)
}
datumCounts = append(datumCounts, c)
}
for idx, c := range frozencounter.FreezeMany(datumCounts) {
data[idx].featureCounts = c
}
counts = frozencounter.NewCounterVector(frozencounter.FreezeMap(rawCounts))
features = data[0].featureCounts.Keys
for label, _ := range counts.Extract() {
labels = append(labels, label)
}
return
}
// Convert a frozen counter back into a counter.Counter.
func (c *Counter) Thaw() *counter.Counter {
t := counter.New(c.Keys.Base)
for s, idx := range c.Keys.Positions {
t.Set(s, c.values[idx])
}
return t
}
// Calculate the label distribution of features given weights, storing the result in out
func (w *maxentWeights) labelDistribution(counts *frozencounter.Counter, weights *frozencounter.CounterVector) *frozencounter.Counter {
out := counter.New(0.0)
for label, featureWeights := range weights.Extract() {
out.Set(label, featureWeights.DotProduct(counts))
}
out.LogNormalize()
return frozencounter.Freeze(out)
}
