Source code for theanolm.backend.probfunctions

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""A module that provides functions for operating with probabilities, and
defines logprob_type.
"""

from decimal import Decimal, getcontext

import numpy
import theano

logprob_type = numpy.dtype(theano.config.floatX).type



def interpolate_linear(logprob1, logprob2, weight1):
    """Performs linear interpolation of two probabilities.

    If a probability is so small that it will be rounded to zero, uses the
    decimal library. If a weight is zero, the corresponding log probability will
    be ignored. Otherwise if the log probability was ``-inf``, multiplication
    would result in a ``nan``.

    :type logprob1: logprob_type
    :param logprob1: logarithm of first input probability

    :type logprob2: logprob_type
    :param logprob2: logarithm of second input probability

    :type weight1: logprob_type
    :param weight1: interpolation weight for the first probability

    :rtype: logprob_type
    :returns: logarithm of the weighted sum of the input probabilities
    """

    prob1 = numpy.exp(numpy.float64(logprob1))
    prob2 = numpy.exp(numpy.float64(logprob2))
    if (prob1 > 0) and (prob2 > 0):
        weight2 = 1.0 - weight1
        prob = weight1 * prob1
        prob += weight2 * prob2
        return logprob_type(numpy.log(prob))
    else:
        # An exp() resulted in an underflow (or -inf logprob). Use the decimal
        # library.
        getcontext().prec = 16
        d_weight1 = Decimal(float(weight1))
        d_weight2 = Decimal(1.0) - d_weight1
        d_logprob1 = Decimal(float(logprob1))
        d_logprob2 = Decimal(float(logprob2))
        d_zero = Decimal(0.0)
        d_prob = d_zero
        if d_weight1 != d_zero:
            d_prob += d_weight1 * d_logprob1.exp()
        if d_weight2 != d_zero:
            d_prob += d_weight2 * d_logprob2.exp()
        return logprob_type(d_prob.ln())




def interpolate_loglinear(logprob1, logprob2, prior1, prior2):
    """Performs log-linear interpolation of two probabilities.

    This is not real log-linear interpolation, because we don't normalize the
    result. Thus the result is not a real probability.

    If a prior is zero, the corresponding log probability will be ignored.
    Otherwise if the log probability was ``-inf``, multiplication would result
    in a ``nan``.

    :type logprob1: logprob_type
    :param logprob1: first input log probability

    :type logprob2: logprob_type
    :param logprob2: second input log probability

    :type prior1: logprob_type
    :param prior1: weight for the first log probability

    :type prior2: logprob_type
    :param prior2: weight for the second log probability

    :rtype: logprob_type
    :returns: weighted sum of the input log probabilities
    """

    result = 0
    if prior1 != 0:
        result += prior1 * logprob1
    if prior2 != 0:
        result += prior2 * logprob2
    assert not numpy.isnan(result)
    return result