BayesianNeuralNetwork.py 文件源码

def __log_likelihood_factor__(self, samples_q, v_noise, X, wb, y):
        # Account for occasions where we're optimizing the latent weighting distributions
        if wb.shape[0] == 1:
            if wb.shape[1] > self.num_latent_params: # Further account
                # Reshape the wb to be a full matrix and build full latent array
                Wb = np.reshape(wb, [-1,self.num_latent_params])
                latent_weights = np.array([Wb[int(X[tt,-1]),:] for tt in range(X.shape[0])])
                outputs = self.__predict__(samples_q, np.hstack([X[:,:-1], latent_weights]))
            else:
                outputs = self.__predict__(samples_q, np.hstack([X, np.tile(wb,(X.shape[0],1))]))
        else:
            outputs = self.__predict__(samples_q, np.hstack([X, wb]))
        return (-0.5*np.log(2*math.pi*v_noise)) - (0.5*(np.tile(np.expand_dims(y,axis=0), (self.num_weight_samples,1,1))-outputs)**2)/v_noise