本文主要是介绍SRGAN loss部分的mindspore代码实现，希望对大家解决编程问题提供一定的参考价值，需要的开发者们随着小编来一起学习吧！
转载地址：https://bbs.huaweicloud.com/forum/thread-137103-1-1.html
作者： 雨丝儿
pytorch代码实现:https://bbs.huaweicloud.com/forum/forum.php?mod=viewthread&tid=137101
接上篇帖子 ，本贴分享mindspore代码的实现：
import mindspore.common.dtype as mstype
import mindspore.nn as nn
import mindspore.ops as ops
from src.vgg19.define import vgg19
from src.loss.Meanshift import MeanShift
对于Discriminator:
class DiscriminatorLoss(nn.Cell):
"""Loss for discriminator"""
def __init__(self, discriminator, generator):
super(DiscriminatorLoss, self).__init__()
self.discriminator = discriminator
self.generator = generator
self.adversarial_criterion = nn.BCELoss()
ones = ops.Ones()
zeros = ops.Zeros()
self.real_lable = ones((16, 1), mstype.float32)
self.fake_lable = zeros((16, 1), mstype.float32)
def construct(self, HR_img, LR_img):
"""dloss"""
hr = HR_img
lr = LR_img
# Generating fake high resolution images from real low resolution images.
sr = self.generator(lr)
# Let the discriminator realize that the sample is real.
real_output = self.discriminator(hr)
d_loss_real = self.adversarial_criterion(real_output, self.real_lable)
# Let the discriminator realize that the sample is false.
fake_output = self.discriminator(sr)
d_loss_fake = self.adversarial_criterion(fake_output, self.fake_lable)
d_loss = d_loss_fake+d_loss_real
return d_loss
class GeneratorLoss(nn.Cell):
"""Loss for generator"""
def __init__(self, discriminator, generator, vgg_ckpt):
super(GeneratorLoss, self).__init__()
self.discriminator = discriminator
self.generator = generator
self.mse_loss = nn.MSELoss()
self.adversarial_criterion = nn.BCELoss()
ones = ops.Ones()
self.real_lable = ones((16, 1), mstype.float32)
self.meanshif = MeanShift()
self.vgg = vgg19(vgg_ckpt)
for p in self.meanshif.get_parameters():
p.requires_grad = False
def construct(self, HR_img, LR_img):
"""gloss"""
# L2loss
hr = HR_img
lr = LR_img
sr = self.generator(lr)
L2_loss = self.mse_loss(sr, hr)
# adversarialloss
fake_output = self.discriminator(sr)
adversarial_loss = self.adversarial_criterion(fake_output, self.real_lable)
# vggloss
hr = (hr+1.0)/2.0
sr = (sr+1.0)/2.0
hr = self.meanshif(hr)
sr = self.meanshif(sr)
hr_feat = self.vgg(hr)
sr_feat = self.vgg(sr)
percep_loss = self.mse_loss(hr_feat, sr_feat)
g_loss = 0.006*percep_loss+1e-3*adversarial_loss+L2_loss
return g_loss
这篇关于SRGAN loss部分的mindspore代码实现的文章就介绍到这儿，希望我们推荐的文章对编程师们有所帮助！
