def test(self):
self.real_A = Variable(self.input_A, volatile=True)
self.fake_B = self.netG.forward(self.real_A)
self.real_B = Variable(self.input_B, volatile=True)
self.pre_filter, self.depth = self.netDepth.forward(self.real_A)
# recover B according to depth
self.fake_B2 = util.reverse_matting(self.real_A, self.depth)
# reconstruct A based on optical model
self.fake_A = util.synthesize_matting(self.fake_B, self.depth)
def forward(self):
self.real_A = Variable(self.input_A)
self.fake_B = self.netG.forward(self.real_A)
self.real_B = Variable(self.input_B)
self.depth = self.netDepth.forward(self.real_A)
# clip with 0.9
self.depth = torch.clamp(self.depth, max=0.9)
# recover B according to depth
self.fake_B2 = util.reverse_matting(self.real_A, self.depth)
# reconstruct A based on optical model
self.fake_A = util.synthesize_matting(self.fake_B, self.depth)
def forward(self):
self.real_A = Variable(self.input_A)
self.fake_B = self.netG.forward(
self.real_A) #GAN生成hazefreeimage fake_B
self.real_B = Variable(self.input_B)
self.depth = self.netDepth.forward(self.real_A) #netdepth输生成透射率图
# clip with 0.9
self.depth = torch.clamp(self.depth, max=0.9)
# recover B according to depth
self.fake_B2 = util.reverse_matting(
self.real_A,
self.depth) #根据透射图depth和默认大气光1.0被hazeimage恢复为hazefree image
# reconstruct A based on optical model
self.fake_A = util.synthesize_matting(
self.fake_B, self.depth) #根据hazefreeimage和透射图depth生成haze image
def test(self):
self.real_A = Variable(self.input_A, volatile=True)
self.fake_B = self.netG.forward(self.real_A)
self.real_B = Variable(self.input_B, volatile=True)
self.depth = self.netDepth.forward(self.real_A)
if not self.opt.which_model_depth == 'aod':
self.depth = (self.depth + 1) / 2.
# recover B according to depth
self.fake_B2 = util.reverse_matting(self.real_A, self.depth)
# reconstruct A based on optical model
self.fake_A = util.synthesize_matting(self.fake_B, self.depth)
# feed haze-free image
self.extra_B = self.netG.forward(self.real_B)
self.extra_depth = self.netDepth.forward(self.real_B)
if not self.opt.which_model_depth == 'aod':
self.extra_depth = (self.extra_depth + 1) / 2. # scale it to [0,1]
self.extra_A = util.synthesize_matting(self.extra_B, self.extra_depth)
def test(self):
self.real_A = Variable(self.input_A, volatile=True)
self.fake_B = self.netG.forward(self.real_A)
self.real_B = Variable(self.input_B, volatile=True)
self.depth = self.netDepth.forward(self.real_A)
if not self.opt.which_model_depth == 'aod':
self.depth = (self.depth + 1) / 2.
# regularize depth by lower bound
real_A = (self.real_A + 1) / 2 # scale it to [0,1]
A = 1. # A is 1 at this moment
LB = 1 - torch.min(real_A / A, dim=1,
keepdim=True)[0] # get lower bould of depth
self.depth_LB = torch.max(self.depth, LB)
# recover B according to depth
self.fake_B2 = util.reverse_matting(self.real_A, self.depth_LB)
# reconstruct A based on optical model
self.fake_A = util.synthesize_matting(self.fake_B, self.depth_LB)
