版权声明：本文为博主原创文章，未经博主允许不得转载。 https://blog.csdn.net/qq_41185868/article/details/82941252
CV之NS之Loss：NS中常用的几种损失简介、使用方法之详细攻略
1、内容损失
# endpoints_dict是上一节提到的损失网络各层的计算结果；content_layers是定义使用哪些层的差距计算损失，默认配置是conv3_3
def content_loss(endpoints_dict, content_layers):
content_loss = 0
for layer in content_layers:
#上一节中把生成的图像、原始图像同时传入损失网络中计算。所以这里先把他们区分开
#我们可以参照函数tf.concat与tf.split的文档理解此处的内容
generated_images, content_images = tf.split(endpoints_dict[layer], 2, 0)
size = tf.size(generated_images)
# 所谓的内容损失，是生成图片generated_images与原始图片激活content_images的L*L距离
content_loss += tf.nn.l2_loss(generated_images - content_images) * 2 / tf.to_float(size) # remain the same as in the paper
return content_loss2、风格损失
# 定义风格损失，style_layers为定义使用哪些层计算风格损失。默认为conv1_2、conv2_2、conv3_3、conv4_3
# style_features_t是利用原始的风格图片计算的层的激活。如在wave模型中是img/wave.jpg计算的激活
def style_loss(endpoints_dict, style_features_t, style_layers):
style_loss = 0
# summary是为TensorBoard服务的
style_loss_summary = {}
for style_gram, layer in zip(style_features_t, style_layers):
# 计算风格损失，只需要计算生成图片generated_imgs与目标风格style_features_t的差距。因此不需要取出content_images
generated_images, _ = tf.split(endpoints_dict[layer], 2, 0)
size = tf.size(generated_images)
# 调用gram函数计算Gram矩阵。风格损失定义为生成图片与目标风格Gram矩阵的L*L的Loss
layer_style_loss = tf.nn.l2_loss(gram(generated_images) - style_gram) * 2 / tf.to_float(size)
style_loss_summary[layer] = layer_style_loss
style_loss += layer_style_loss
return style_loss, style_loss_summary3、定义总损失
"""Build Losses"""
# 定义内容损失
content_loss = losses.content_loss(endpoints_dict, FLAGS.content_layers)
# 定义风格损失
style_loss, style_loss_summary = losses.style_loss(endpoints_dict, style_features_t, FLAGS.style_layers)
# 定义tv损失，该损失在实际训练中并没有被用到，因为在训练时都采用tv_weight=0
tv_loss = losses.total_variation_loss(generated) # use the unprocessed image
# 总损失是这些损失的加权和，最后利用总损失优化图像生成网络即可
loss = FLAGS.style_weight * style_loss + FLAGS.content_weight * content_loss + FLAGS.tv_weight * tv_loss