TensorFlow KLDivergence loss为负

把

kl_compute = tf.keras.losses.KLDivergence(reduction=losses_utils.ReductionV2.NONE, 
    name='kullback_leibler_divergence')
                                                          
loss = tf.reduce_mean(kl_compute(labels, tf.nn.softmax(logits)))

改为

kl_compute = tf.keras.losses.KLDivergence(reduction=losses_utils.ReductionV2.NONE, 
    name='kullback_leibler_divergence')
                                                          
loss = tf.reduce_mean(kl_compute(labels * 10, tf.nn.softmax(logits)))

即labels改为labels * 10