def cross_entropy_logits(self, yhat, M):
if self.smoothing:
n_out = M + self.n_sample
return T.cast(
T.mean((1.0 -
(n_out /
(n_out - 1)) * self.smoothing) * gpu_diag_wide(yhat) +
(self.smoothing / (n_out - 1)) * T.sum(yhat, axis=1)),
theano.config.floatX)
else:
return T.cast(T.mean(gpu_diag_wide(yhat)), theano.config.floatX)
def top1(self, yhat, M):
ydiag = gpu_diag_wide(yhat).dimshuffle((0, 'x'))
return T.cast(
T.mean(
T.mean(T.nnet.sigmoid(-ydiag + yhat) + T.nnet.sigmoid(yhat**2),
axis=1) - T.nnet.sigmoid(ydiag**2) /
(M + self.n_sample)), theano.config.floatX)
def bpr_max(self, yhat, M):
softmax_scores = self.softmax_neg(yhat)
return T.cast(
T.mean(-T.log(
T.sum(T.nnet.sigmoid(
gpu_diag_wide(yhat).dimshuffle(
(0, 'x')) - yhat) * softmax_scores,
axis=1) + 1e-24) +
self.bpreg * T.sum((yhat**2) * softmax_scores, axis=1)),
theano.config.floatX)
def top1_max(self, yhat, M):
softmax_scores = self.softmax_neg(yhat)
y = softmax_scores * (T.nnet.sigmoid(-gpu_diag_wide(yhat).dimshuffle(
(0, 'x')) + yhat) + T.nnet.sigmoid(yhat**2))
return T.cast(T.mean(T.sum(y, axis=1)), theano.config.floatX)
def bpr(self, yhat, M):
return T.cast(
T.mean(-T.log(
T.nnet.sigmoid(
gpu_diag_wide(yhat).dimshuffle((0, 'x')) - yhat))),
theano.config.floatX)
