def psnr(pred, truth):
"""
https://qiita.com/yoya/items/510043d836c9f2f0fe2f
"""
batch_size = len(pred)
mse = F.sum((pred - truth).reshape(batch_size, -1)**2, axis=1)
mse = F.clip(mse, 1e-8, 1e+8)
max_i = F.max(truth.reshape(batch_size, -1), axis=1)
max_i = F.clip(max_i, 1e-8, 1e+8)
return 20 * F.log10(max_i) - 10 * F.log10(mse)
def lf(x, t):
h1, h2, h3, h4 = self.encode(x)
y = self.decode(h1, h2, h3, h4)
mse_loss = F.mean_squared_error(y, t)
with chainer.using_config('train',
False), chainer.no_backprop_mode():
y_content = self.vgg16(y)
t_content = self.vgg16(t)
cont_loss = 0.0
k = [0.9, 0.7, 0.5, 0.3, 0.1]
for i in range(len(y_content)):
cont_loss += k[i] * F.mean_squared_error(
y_content[i], t_content[i])
self.mse_loss = mse_loss
self.cont_loss = cont_loss
self.loss = mse_loss + C * self.cont_loss
self.psnr = 10 * F.log10(1.0 / mse_loss)
chainer.report(
{
'mse_loss': self.mse_loss,
'cont_loss': self.cont_loss,
'loss': self.loss,
'psnr': self.psnr
},
observer=self)
return self.loss
def __call__(self, x, t):
self.y = None
self.loss = None
self.psnr = None
self.y = self.generator(x)
loss_mse = F.mean_squared_error(self.y, t)
with chainer.using_config('train', False):
with chainer.using_config('enable_backprop', False):
y_cont = self.vgg16(self.y)
t_cont = self.vgg16(t)
loss_cont = None
for i in range(len(y_cont)):
if loss_cont is None:
loss_cont = F.mean_squared_error(y_cont[i], t_cont[i])
else:
loss_cont += F.mean_squared_error(y_cont[i], t_cont[i])
self.loss = 10 * loss_cont + loss_mse
self.psnr = 10 * F.log10(1.0 / loss_mse)
reporter.report(
{
'loss': self.loss,
'loss_cont': loss_cont,
'loss_mse': loss_mse,
'PSNR': self.psnr
}, self)
return self.loss
def __call__(self, x, t):
self.y = None
self.loss = None
self.psnr = None
self.y = self.generator(x)
self.loss = F.mean_squared_error(self.y, t)
self.psnr = 10 * F.log10(1.0 / self.loss)
reporter.report({'loss': self.loss, 'PSNR': self.psnr}, self)
return self.loss
def lf(x, t):
y = self(x)
loss = F.mean_squared_error(y, t)
self.loss = loss
self.psnr = 10 * F.log10(1.0 / loss)
chainer.report({
'loss': self.loss,
'psnr': self.psnr
},
observer=self)
return self.loss
def lf(x, t):
h1, h2, h3, h4 = self.encode(x)
y = self.decode(h1, h2, h3, h4)
loss = F.mean_squared_error(y, t)
self.loss = loss
self.psnr = 10 * F.log10(1.0 / loss)
chainer.report({
'loss': self.loss,
'psnr': self.psnr
},
observer=self)
return self.loss
def log10(self, x):
return F.log10(x)
def sid(self, y, t):
err = F.absolute(F.sum(y * F.log10((y + self.eps) / (t + self.eps))) + \
F.sum(t * F.log10((t + self.eps) / (y + self.eps))))
return err * 0.041 # = err * 2** 16 / (14 * 240 * 480)
