Stable binary cross-entropy, operating on logit predictions instead of relying on Theano to correctly optimize the sigmoid output.

bcefl.py

from theano import tensor

def binary_crossentropy_from_logits(logits, targets):
    """Binary cross-entropy computed from model logits.
    Parameters
    ----------
    predictions : TensorVariable
        The unnormalized log probabilities of a probabilistic binary
        classifier.
    targets : TensorVariable
        The targets for the classifier in [0, 1].
    Returns
    -------
    TensorVariable
        The log probability of each prediction under the dirac
        density specified by the corresponding target.
    """
    a, t = logits, targets
    return t * tensor.nnet.softplus(-a) + (1 - t) * tensor.nnet.softplus(a)