def _loss(self, y_true, y_pred):
"""Loss function for keras training.
We assume a model of the form P(x)=exp(-E(x))P_0(x)/Z.
We minimize the neg-loglikelihood: <-logP> = log(Z) - <-E>.
Normalization of P gives Z=<exp(-E)>_{P_0}.
We fix the gauge by adding the constraint (Z-1)**2 to the likelihood.
"""
data= ksum((-y_pred)*(1.-y_true))/ksum(1.-y_true)
gen= klog(ksum(kexp(-y_pred)*y_true))-klog(ksum(y_true))
reg= kexp(gen)-1.
return gen-data+self.gamma*reg*reg
def _likelihood(self, y_true, y_pred):
"""Loss function for keras training.
We assume a model of the form P(x)=exp(-E(x))P_0(x)/Z.
We minimize the neg-loglikelihood: <-logP> = log(Z) - <-E>.
Normalization of P gives Z=<exp(-E)>_{P_0}.z
"""
data= ksum((-y_pred)*(1.-y_true))/ksum(1.-y_true)
gen= klog(ksum(kexp(-y_pred)*y_true))-klog(ksum(y_true))
return gen-data