loss.py 文件源码

def eval_reg(log_sigma2, W):
    # Approximates the negative of the KL-divergence according to eqn 14.
    # This is a key part of the loss function (see eqn 3).
    k1, k2, k3 = 0.63576, 1.8732, 1.48695
    C = -k1
    log_alpha = clip(log_sigma2 - tf.log(W**2))
    mdkl = k1 * tf.nn.sigmoid(k2 + k3 * log_alpha) - 0.5 * tf.log1p(tf.exp(-log_alpha)) + C
    return -tf.reduce_sum(mdkl)