alextp

def hash_collapsed_likelihood(tprior, gmean, gvar, tcount, s1):
    "Assumes a single hyperparameter"
    p0 = np.log(gamma_pdf(tprior, gmean, gvar))
    s0 = len(tcount)
    p0 += s0*(gammaln(tprior*s1)-sum(gammaln(tprior) for i in xrange(s1)))
    for d in xrange(s0):
        for i,t in tcount[d].items():
            p0 += gammaln(t + tprior)
        p0 += (s1-len(tcount[d]))*gammaln(tprior)
        p0 -= gammaln(sum(tcount[d]) + s1*tprior)

alextp

def single_collapsed_likelihood(tprior, gmean, gvar, tcount):
    "Assumes tprior is a scalar"
    p0 = np.log(gamma_pdf(tprior, gmean, gvar))
    p0 += tcount.shape[0]*(gammaln(tprior*tcount.shape[1])-gammaln(tprior)*tcount.shape[1])
    for d in xrange(tcount.shape[0]):
        for i,t in enumerate(tcount[d]):
            p0 += gammaln(t + tprior)
        p0 -= gammaln(sum(tcount[d]) + tprior*tcount.shape[1])
    return p0

alextp

def collapsed_likelihood(tprior, gmean, gvar, tcount):
   p0 = sum(np.log(gamma_pdf(tp, gmean, gvar)) for tp in tprior)
   p0 += tcount.shape[0]*(gammaln(sum(tprior))-sum(gammaln(tp) for tp in tprior))
   for d in xrange(tcount.shape[0]):
       for i,t in enumerate(tcount[d]):
           p0 += gammaln(t + tprior[i])
       p0 -= gammaln(sum(tcount[d]) + sum(tprior))
   return p0

alextp

class OnlineLearner(object):
    def __init__(self, **kwargs):
        self.last_misses = 0.
        self.iratio = 0.
        self.it = 1.
        self.l = kwargs["l"]
        self.max_ratio = -np.inf
        self.threshold = 500.

    def hinge_loss(self, vector, cls, weight):