d57bfbec8e170a6c8517d60b585393b0f3b3c732,gpflow/models/svgp.py,SVGP,log_likelihood,#SVGP#,137

Before Change

        kl = self.prior_kl()
        f_mean, f_var = self.predict_f(X)
        var_exp = self.likelihood.variational_expectations(f_mean, f_var, Y)
        scale = tf.cast(self.num_data, kl.dtype) / tf.cast(tf.shape(X)[0], kl.dtype)
        return tf.reduce_sum(var_exp) * scale - kl

    def predict_f(self, Xnew, full_cov=False, full_output_cov=False) -> tf.Tensor:
        q_mu = self.q_mu()

After Change

        var_exp = self.likelihood.variational_expectations(f_mean, f_var, Y)
        scale = tf.cast(1.0, kl.data)
        if self.num_data is not None:
            num_data = tf.cast(self.num_data, kl.dtype)
            minibatch_size = tf.cast(tf.shape(X)[0], kl.dtype)
            scale = num_data / minibatch_size
        return tf.reduce_sum(var_exp) * scale - kl

Instances