def crossentropy_loss(self,total_input,targets,grad=False):
log_preds = nu.log_sigmoid(total_input)
log_one_minus_preds = nu.log_sigmoid(-total_input)
preds = np.exp(log_preds)
if (not grad):
loss = -np.sum(targets*log_preds + (1-targets)*log_one_minus_preds,1).mean(0)
return loss
else:
dloss = (preds - targets) / targets.shape[0]
return dloss
def nonlin(self,total_input):
return -nu.log_sigmoid(-total_input)
