BernoulliNADE.py 文件源码

def sym_logdensity(self, x):
        """ x is a matrix of column datapoints (VxB) V = n_visible, B = batch size """
        def density_given_previous_a_and_x(x, w, v, b, activations_factor, p_prev, a_prev, x_prev):
            a = a_prev + T.dot(T.shape_padright(x_prev, 1), T.shape_padleft(w, 1))
            h = self.nonlinearity(a * activations_factor)  # BxH
            t = T.dot(h, v) + b
            p_xi_is_one = T.nnet.sigmoid(t) * constantX(0.9999) + constantX(0.0001 * 0.5)  # Make logistic regression more robust by having the sigmoid saturate at 0.00005 and 0.99995
            p = p_prev + x * T.log(p_xi_is_one) + (1 - x) * T.log(1 - p_xi_is_one)
            return (p, a, x)
        # First element is different (it is predicted from the bias only)
        a0 = T.zeros_like(T.dot(x.T, self.W))  # BxH
        p0 = T.zeros_like(x[0])
        x0 = T.ones_like(x[0])
        ([ps, _, _], updates) = theano.scan(density_given_previous_a_and_x,
                                            sequences=[x, self.W, self.V, self.b, self.activation_rescaling],
                                            outputs_info=[p0, a0, x0])
        return (ps[-1], updates)