def max_likelihood_qa(data, nvals, qa, approximate=False, smooth=0.1):
nc = qa.max() + 1
sil = np.zeros(nc)
for i in range(nc):
di = data[qa==i].astype(np.int)
total = float(len(di))
si = 0.0
for j in range(di.shape[1]):
bc = np.bincount(di[:,j])
for c in bc:
if c == 0:
continue
si += c*log(c / total)
si += log(total / len(data)) # cluster prior
sil[i] = si
s = sil.sum() if approximate else logsumexp(sil)
return s
def penalized_ll(nbcs, data):
ll = 0.0
cluster_priors = [log(1.0 * c.size / len(data)) for c in nbcs]
for inst in data:
vals = [cluster_priors[i] + nbcs[i].fast_ll(inst, float('-inf')) for i in range(len(nbcs))]
ll += logsumexp( vals )
ll -= knobs.cluster_penalty * len(nbcs) * data.shape[1]
return ll
def ll(self, inst, best_cll=float('-inf')):
sstats = self.sstats
l = 0.0
for i in xrange(len(self.nvals)):
v = inst[i]
w = sstats[i][v]
l += log((w + self.smoo) / (self.size + self.nvals[i]*self.smoo))
if l < best_cll:
return l
return l
def nbmix_model(data, nvals, qa, smooth=0.1):
data = data.astype(np.int, copy=False)
nc = qa.max() + 1
n = data.shape[1]
m = data.shape[0]
# compute params for NB models
lprobs = float('-inf')*np.ones( (n, nc, max(nvals)) )
priors = np.zeros(nc)
for i in range(nc):
di = data[qa==i]
di_size = float(len(di))
priors[i] = log(di_size / m)
for j in range(n):
bc = np.bincount(di[:,j], minlength=nvals[j])
for k in range(nvals[j]):
c = bc[k]
if c == 0:
continue
lprobs[j,i,k] = log((c + smooth) / (di_size + smooth*nvals[j]))
return lprobs, priors
def new_ll(nvals, smooth=0.1):
return sum( log((1. + smooth) / (1. + v * smooth)) for v in nvals )
