def update_netd(self):
"""
Update D network: Ladv = |f(real) - f(fake)|_2
"""
##
# Feature Matching.
self.netd.zero_grad()
# --
# Train with real
self.label.data.resize_(self.input.size(0)).fill_(self.real_label)
self.out_d_real, self.feat_real = self.netd(self.input)
# self.err_d_real = self.bce_criterion(self.out_d_real,self.label)
# Train with fake
self.label.data.resize_(self.input.size(0)).fill_(self.fake_label)
self.fake, self.latent_i, = self.netg(self.input)
self.out_d_fake, self.feat_fake = self.netd(self.fake.detach())
# self.err_d_fake = self.bce_criterion(self.out_d_fake, self.label)
# --
# self.err_d = self.err_d_real + self.err_d_fake
self.err_d = l2_loss(self.feat_real, self.feat_fake)
# self.err_d = self.err_d_fake + self.err_d_l2
# self.err_d_real = self.err_d
# self.err_d_fake = self.err_d
self.err_d.backward()
self.optimizer_d.step()
def update_netd(self):
"""
Update D network: Ladv = |f(real) - f(fake)|_2
"""
##
# Feature Matching.
self.netd.zero_grad()
# --
# Train with real
self.out_d_real, self.feat_real = self.netd(self.input)
# --
# Train with fake
self.fake, self.latent_i, self.latent_o = self.netg(self.input)
self.out_d_fake, self.feat_fake = self.netd(self.fake.detach())
# --
self.err_d = l2_loss(self.feat_real, self.feat_fake)
self.err_d.backward()
self.optimizer_d.step()
def update_netd(self):
"""
Update D network: Ladv = |f(real) - f(fake)|_2
"""
##
# Feature Matching.
self.netd.zero_grad()
# --
# Train with real
with torch.no_grad():
self.label.resize_(self.opt.batchsize).fill_(self.real_label)
self.out_d_real, self.feat_real = self.netd(self.input)
# --
# Train with fake
with torch.no_grad():
self.label.resize_(self.opt.batchsize).fill_(self.fake_label)
self.fake, self.latent_i, self.latent_o = self.netg(self.input)
self.out_d_fake, self.feat_fake = self.netd(self.fake.detach())
# --
self.err_d = l2_loss(self.feat_real, self.feat_fake)
self.err_d_real = self.err_d
self.err_d_fake = self.err_d
self.err_d.backward()
self.optimizer_d.step()
