图像分割loss对比

1.binary_crossentropy

 # Loss最终用binary_crossentropy loss，sigmoid激活后放到一个logit的loss里
 loss = tf.reduce_mean(tf.keras.losses.binary_crossentropy(labels, pred))

2.sigmoid_cross_entropy_with_logits 不行

 # #  第一种loss， sigmoid_cross_entropy_with_logits输入为未激活的值
 # loss_h = tf.reduce_mean(
 #     tf.nn.sigmoid_cross_entropy_with_logits(logits=pred[:, :, :, 0:1], labels=labels[:, :, :, 0:1]))
 # loss_s = tf.reduce_mean(
 #     tf.nn.sigmoid_cross_entropy_with_logits(logits=pred[:, :, :, 1:2], labels=labels[:, :, :, 1:2]))
 # loss = loss_h + loss_s

3.sigmoid_cross_entropy 不行

# # 第三种loss 与第一种一模一样
# loss = tf.compat.v1.losses.sigmoid_cross_entropy(labels, pred)