Python tensorboard.summary.scalar() Examples
The following are 19
code examples of tensorboard.summary.scalar().
You can vote up the ones you like or vote down the ones you don't like,
and go to the original project or source file by following the links above each example.
You may also want to check out all available functions/classes of the module
tensorboard.summary
, or try the search function
.
.
Example #1
| Source File: test_logging.py From tensorboard with Apache License 2.0 | 6 votes |
|
def test_event_logging():
logdir = './experiment/'
summary_writer = FileWriter(logdir)
scalar_value = 1.0
s = scalar('test_scalar', scalar_value)
summary_writer.add_summary(s, global_step=1)
summary_writer.close()
assert os.path.isdir(logdir)
assert len(os.listdir(logdir)) == 1
summary_writer = FileWriter(logdir)
scalar_value = 1.0
s = scalar('test_scalar', scalar_value)
summary_writer.add_summary(s, global_step=1)
summary_writer.close()
assert os.path.isdir(logdir)
assert len(os.listdir(logdir)) == 2
# clean up.
shutil.rmtree(logdir)
Example #2
| Source File: trainer.py From Recipe2ImageGAN with MIT License | 5 votes |
|
def train_Dnet(self, idx, count):
flag = count % 100
batch_size = self.real_imgs[0].size(0)
criterion = self.criterion
netD, optD = self.netsD[idx], self.optimizersD[idx]
real_imgs = self.real_imgs[idx]
fake_imgs = self.fake_imgs[idx]
real_labels = self.real_labels[:batch_size]
fake_labels = self.fake_labels[:batch_size]
#
netD.zero_grad()
#
real_logits = netD(real_imgs)
fake_logits = netD(fake_imgs.detach())
#
errD_real = criterion(real_logits[0], real_labels)
errD_fake = criterion(fake_logits[0], fake_labels)
#
errD = errD_real + errD_fake
errD.backward()
# update parameters
optD.step()
# log
if flag == 0:
summary_D = summary.scalar('D_loss%d' % idx, errD.data[0])
self.summary_writer.add_summary(summary_D, count)
return errD
Example #3
| Source File: trainer.py From StackGAN-v2 with MIT License | 5 votes |
|
def train_Dnet(self, idx, count):
flag = count % 100
batch_size = self.real_imgs[0].size(0)
criterion = self.criterion
netD, optD = self.netsD[idx], self.optimizersD[idx]
real_imgs = self.real_imgs[idx]
fake_imgs = self.fake_imgs[idx]
real_labels = self.real_labels[:batch_size]
fake_labels = self.fake_labels[:batch_size]
#
netD.zero_grad()
#
real_logits = netD(real_imgs)
fake_logits = netD(fake_imgs.detach())
#
errD_real = criterion(real_logits[0], real_labels)
errD_fake = criterion(fake_logits[0], fake_labels)
#
errD = errD_real + errD_fake
errD.backward()
# update parameters
optD.step()
# log
if flag == 0:
summary_D = summary.scalar('D_loss%d' % idx, errD.data[0])
self.summary_writer.add_summary(summary_D, count)
return errD
Example #4
| Source File: test_summary_writer.py From tensorboard with Apache License 2.0 | 5 votes |
|
def test_scalar_summary():
scalar_value = 1.0
s = scalar('test_scalar', scalar_value)
values = s.value
assert len(values) == 1
assert values[0].tag == 'test_scalar'
assert values[0].simple_value == 1.0
byte_str = s.SerializeToString()
s_recovered = summary_pb2.Summary()
s_recovered.ParseFromString(byte_str)
assert values[0].tag == s_recovered.value[0].tag
assert values[0].simple_value == s_recovered.value[0].simple_value
Example #5
| Source File: eval_trainer.py From Recipe2ImageGAN with MIT License | 5 votes |
|
def train_Dnet(self, idx, count):
flag = count % 100
batch_size = self.real_imgs[0].size(0)
criterion = self.criterion
netD, optD = self.netsD[idx], self.optimizersD[idx]
real_imgs = self.real_imgs[idx]
fake_imgs = self.fake_imgs[idx]
real_labels = self.real_labels[:batch_size]
fake_labels = self.fake_labels[:batch_size]
#
netD.zero_grad()
#
real_logits = netD(real_imgs)
fake_logits = netD(fake_imgs.detach())
#
errD_real = criterion(real_logits[0], real_labels)
errD_fake = criterion(fake_logits[0], fake_labels)
#
errD = errD_real + errD_fake
errD.backward()
# update parameters
optD.step()
# log
if flag == 0:
summary_D = summary.scalar('D_loss%d' % idx, errD.data[0])
self.summary_writer.add_summary(summary_D, count)
return errD
Example #6
| Source File: test_summary_writer.py From tensorboard with Apache License 2.0 | 5 votes |
|
def test_log_scalar_summary():
logdir = './experiment/scalar'
writer = SummaryWriter(logdir)
for i in range(10):
writer.add_scalar('test_scalar', i+1)
writer.close()
Example #7
| Source File: blah.py From Recipe2ImageGAN with MIT License | 5 votes |
|
def train_Dnet(self, idx, count):
flag = count % 100
batch_size = self.real_imgs[0].size(0)
criterion = self.criterion
netD, optD = self.netsD[idx], self.optimizersD[idx]
real_imgs = self.real_imgs[idx]
fake_imgs = self.fake_imgs[idx]
real_labels = self.real_labels[:batch_size]
fake_labels = self.fake_labels[:batch_size]
#
netD.zero_grad()
#
real_logits = netD(real_imgs)
fake_logits = netD(fake_imgs.detach())
#
errD_real = criterion(real_logits[0], real_labels)
errD_fake = criterion(fake_logits[0], fake_labels)
#
errD = errD_real + errD_fake
errD.backward()
# update parameters
optD.step()
# log
if flag == 0:
summary_D = summary.scalar('D_loss%d' % idx, errD.data[0])
self.summary_writer.add_summary(summary_D, count)
return errD
Example #8
| Source File: test_scalar.py From tensorboard with Apache License 2.0 | 5 votes |
|
def test_log_scalar_summary():
logdir = './experiment/scalar'
writer = FileWriter(logdir)
for i in range(10):
s = scalar('scalar', i)
writer.add_summary(s, i+1)
writer.flush()
writer.close()
Example #9
| Source File: test_scalar.py From tensorboard with Apache License 2.0 | 5 votes |
|
def test_scalar_summary():
scalar_value = 1.0
s = scalar('test_scalar', scalar_value)
values = s.value
assert len(values) == 1
assert values[0].tag == 'test_scalar'
assert values[0].simple_value == 1.0
byte_str = s.SerializeToString()
s_recovered = summary_pb2.Summary()
s_recovered.ParseFromString(byte_str)
assert values[0].tag == s_recovered.value[0].tag
assert values[0].simple_value == s_recovered.value[0].simple_value
Example #10
| Source File: eval_trainer.py From Recipe2ImageGAN with MIT License | 4 votes |
|
def train_Dnet(self, idx, count):
flag = count % 100
batch_size = self.real_imgs[0].size(0)
criterion, mu = self.criterion, self.mu
netD, optD = self.netsD[idx], self.optimizersD[idx]
real_imgs = self.real_imgs[idx]
wrong_imgs = self.wrong_imgs[idx]
fake_imgs = self.fake_imgs[idx]
#
netD.zero_grad()
# Forward
real_labels = self.real_labels[:batch_size]
fake_labels = self.fake_labels[:batch_size]
# for real
real_logits = netD(real_imgs, mu.detach())
wrong_logits = netD(wrong_imgs, mu.detach())
fake_logits = netD(fake_imgs.detach(), mu.detach())
#
errD_real = criterion(real_logits[0], real_labels)
errD_wrong = criterion(wrong_logits[0], fake_labels)
errD_fake = criterion(fake_logits[0], fake_labels)
if len(real_logits) > 1 and cfg.TRAIN.COEFF.UNCOND_LOSS > 0:
errD_real_uncond = cfg.TRAIN.COEFF.UNCOND_LOSS * \
criterion(real_logits[1], real_labels)
errD_wrong_uncond = cfg.TRAIN.COEFF.UNCOND_LOSS * \
criterion(wrong_logits[1], real_labels)
errD_fake_uncond = cfg.TRAIN.COEFF.UNCOND_LOSS * \
criterion(fake_logits[1], fake_labels)
#
errD_real = errD_real + errD_real_uncond
errD_wrong = errD_wrong + errD_wrong_uncond
errD_fake = errD_fake + errD_fake_uncond
#
errD = errD_real + errD_wrong + errD_fake
else:
errD = errD_real + 0.5 * (errD_wrong + errD_fake)
# backward
errD.backward()
# update parameters
optD.step()
# log
if flag == 0:
summary_D = summary.scalar('D_loss%d' % idx, errD.data[0])
self.summary_writer.add_summary(summary_D, count)
return errD
Example #11
| Source File: trainer.py From StackGAN-v2 with MIT License | 4 votes |
|
def train_Gnet(self, count):
self.netG.zero_grad()
errG_total = 0
flag = count % 100
batch_size = self.real_imgs[0].size(0)
criterion, mu, logvar = self.criterion, self.mu, self.logvar
real_labels = self.real_labels[:batch_size]
for i in range(self.num_Ds):
outputs = self.netsD[i](self.fake_imgs[i], mu)
errG = criterion(outputs[0], real_labels)
if len(outputs) > 1 and cfg.TRAIN.COEFF.UNCOND_LOSS > 0:
errG_patch = cfg.TRAIN.COEFF.UNCOND_LOSS *\
criterion(outputs[1], real_labels)
errG = errG + errG_patch
errG_total = errG_total + errG
if flag == 0:
summary_D = summary.scalar('G_loss%d' % i, errG.data[0])
self.summary_writer.add_summary(summary_D, count)
# Compute color consistency losses
if cfg.TRAIN.COEFF.COLOR_LOSS > 0:
if self.num_Ds > 1:
mu1, covariance1 = compute_mean_covariance(self.fake_imgs[-1])
mu2, covariance2 = \
compute_mean_covariance(self.fake_imgs[-2].detach())
like_mu2 = cfg.TRAIN.COEFF.COLOR_LOSS * nn.MSELoss()(mu1, mu2)
like_cov2 = cfg.TRAIN.COEFF.COLOR_LOSS * 5 * \
nn.MSELoss()(covariance1, covariance2)
errG_total = errG_total + like_mu2 + like_cov2
if flag == 0:
sum_mu = summary.scalar('G_like_mu2', like_mu2.data[0])
self.summary_writer.add_summary(sum_mu, count)
sum_cov = summary.scalar('G_like_cov2', like_cov2.data[0])
self.summary_writer.add_summary(sum_cov, count)
if self.num_Ds > 2:
mu1, covariance1 = compute_mean_covariance(self.fake_imgs[-2])
mu2, covariance2 = \
compute_mean_covariance(self.fake_imgs[-3].detach())
like_mu1 = cfg.TRAIN.COEFF.COLOR_LOSS * nn.MSELoss()(mu1, mu2)
like_cov1 = cfg.TRAIN.COEFF.COLOR_LOSS * 5 * \
nn.MSELoss()(covariance1, covariance2)
errG_total = errG_total + like_mu1 + like_cov1
if flag == 0:
sum_mu = summary.scalar('G_like_mu1', like_mu1.data[0])
self.summary_writer.add_summary(sum_mu, count)
sum_cov = summary.scalar('G_like_cov1', like_cov1.data[0])
self.summary_writer.add_summary(sum_cov, count)
kl_loss = KL_loss(mu, logvar) * cfg.TRAIN.COEFF.KL
errG_total = errG_total + kl_loss
errG_total.backward()
self.optimizerG.step()
return kl_loss, errG_total
Example #12
| Source File: trainer.py From StackGAN-v2 with MIT License | 4 votes |
|
def train_Dnet(self, idx, count):
flag = count % 100
batch_size = self.real_imgs[0].size(0)
criterion, mu = self.criterion, self.mu
netD, optD = self.netsD[idx], self.optimizersD[idx]
real_imgs = self.real_imgs[idx]
wrong_imgs = self.wrong_imgs[idx]
fake_imgs = self.fake_imgs[idx]
#
netD.zero_grad()
# Forward
real_labels = self.real_labels[:batch_size]
fake_labels = self.fake_labels[:batch_size]
# for real
real_logits = netD(real_imgs, mu.detach())
wrong_logits = netD(wrong_imgs, mu.detach())
fake_logits = netD(fake_imgs.detach(), mu.detach())
#
errD_real = criterion(real_logits[0], real_labels)
errD_wrong = criterion(wrong_logits[0], fake_labels)
errD_fake = criterion(fake_logits[0], fake_labels)
if len(real_logits) > 1 and cfg.TRAIN.COEFF.UNCOND_LOSS > 0:
errD_real_uncond = cfg.TRAIN.COEFF.UNCOND_LOSS * \
criterion(real_logits[1], real_labels)
errD_wrong_uncond = cfg.TRAIN.COEFF.UNCOND_LOSS * \
criterion(wrong_logits[1], real_labels)
errD_fake_uncond = cfg.TRAIN.COEFF.UNCOND_LOSS * \
criterion(fake_logits[1], fake_labels)
#
errD_real = errD_real + errD_real_uncond
errD_wrong = errD_wrong + errD_wrong_uncond
errD_fake = errD_fake + errD_fake_uncond
#
errD = errD_real + errD_wrong + errD_fake
else:
errD = errD_real + 0.5 * (errD_wrong + errD_fake)
# backward
errD.backward()
# update parameters
optD.step()
# log
if flag == 0:
summary_D = summary.scalar('D_loss%d' % idx, errD.data[0])
self.summary_writer.add_summary(summary_D, count)
return errD
Example #13
| Source File: trainer.py From StackGAN-v2 with MIT License | 4 votes |
|
def train_Gnet(self, count):
self.netG.zero_grad()
errG_total = 0
flag = count % 100
batch_size = self.real_imgs[0].size(0)
criterion = self.criterion
real_labels = self.real_labels[:batch_size]
for i in range(self.num_Ds):
netD = self.netsD[i]
outputs = netD(self.fake_imgs[i])
errG = criterion(outputs[0], real_labels)
# errG = self.stage_coeff[i] * errG
errG_total = errG_total + errG
if flag == 0:
summary_G = summary.scalar('G_loss%d' % i, errG.data[0])
self.summary_writer.add_summary(summary_G, count)
# Compute color preserve losses
if cfg.TRAIN.COEFF.COLOR_LOSS > 0:
if self.num_Ds > 1:
mu1, covariance1 = compute_mean_covariance(self.fake_imgs[-1])
mu2, covariance2 = \
compute_mean_covariance(self.fake_imgs[-2].detach())
like_mu2 = cfg.TRAIN.COEFF.COLOR_LOSS * nn.MSELoss()(mu1, mu2)
like_cov2 = cfg.TRAIN.COEFF.COLOR_LOSS * 5 * \
nn.MSELoss()(covariance1, covariance2)
errG_total = errG_total + like_mu2 + like_cov2
if self.num_Ds > 2:
mu1, covariance1 = compute_mean_covariance(self.fake_imgs[-2])
mu2, covariance2 = \
compute_mean_covariance(self.fake_imgs[-3].detach())
like_mu1 = cfg.TRAIN.COEFF.COLOR_LOSS * nn.MSELoss()(mu1, mu2)
like_cov1 = cfg.TRAIN.COEFF.COLOR_LOSS * 5 * \
nn.MSELoss()(covariance1, covariance2)
errG_total = errG_total + like_mu1 + like_cov1
if flag == 0:
sum_mu = summary.scalar('G_like_mu2', like_mu2.data[0])
self.summary_writer.add_summary(sum_mu, count)
sum_cov = summary.scalar('G_like_cov2', like_cov2.data[0])
self.summary_writer.add_summary(sum_cov, count)
if self.num_Ds > 2:
sum_mu = summary.scalar('G_like_mu1', like_mu1.data[0])
self.summary_writer.add_summary(sum_mu, count)
sum_cov = summary.scalar('G_like_cov1', like_cov1.data[0])
self.summary_writer.add_summary(sum_cov, count)
errG_total.backward()
self.optimizerG.step()
return errG_total
Example #14
| Source File: eval_trainer.py From Recipe2ImageGAN with MIT License | 4 votes |
|
def train_Gnet(self, count):
self.netG.zero_grad()
errG_total = 0
flag = count % 100
batch_size = self.real_imgs[0].size(0)
criterion = self.criterion
real_labels = self.real_labels[:batch_size]
for i in range(self.num_Ds):
netD = self.netsD[i]
outputs = netD(self.fake_imgs[i])
errG = criterion(outputs[0], real_labels)
# errG = self.stage_coeff[i] * errG
errG_total = errG_total + errG
if flag == 0:
summary_G = summary.scalar('G_loss%d' % i, errG.data[0])
self.summary_writer.add_summary(summary_G, count)
# Compute color preserve losses
if cfg.TRAIN.COEFF.COLOR_LOSS > 0:
if self.num_Ds > 1:
mu1, covariance1 = compute_mean_covariance(self.fake_imgs[-1])
mu2, covariance2 = \
compute_mean_covariance(self.fake_imgs[-2].detach())
like_mu2 = cfg.TRAIN.COEFF.COLOR_LOSS * nn.MSELoss()(mu1, mu2)
like_cov2 = cfg.TRAIN.COEFF.COLOR_LOSS * 5 * \
nn.MSELoss()(covariance1, covariance2)
errG_total = errG_total + like_mu2 + like_cov2
if self.num_Ds > 2:
mu1, covariance1 = compute_mean_covariance(self.fake_imgs[-2])
mu2, covariance2 = \
compute_mean_covariance(self.fake_imgs[-3].detach())
like_mu1 = cfg.TRAIN.COEFF.COLOR_LOSS * nn.MSELoss()(mu1, mu2)
like_cov1 = cfg.TRAIN.COEFF.COLOR_LOSS * 5 * \
nn.MSELoss()(covariance1, covariance2)
errG_total = errG_total + like_mu1 + like_cov1
if flag == 0:
sum_mu = summary.scalar('G_like_mu2', like_mu2.data[0])
self.summary_writer.add_summary(sum_mu, count)
sum_cov = summary.scalar('G_like_cov2', like_cov2.data[0])
self.summary_writer.add_summary(sum_cov, count)
if self.num_Ds > 2:
sum_mu = summary.scalar('G_like_mu1', like_mu1.data[0])
self.summary_writer.add_summary(sum_mu, count)
sum_cov = summary.scalar('G_like_cov1', like_cov1.data[0])
self.summary_writer.add_summary(sum_cov, count)
errG_total.backward()
self.optimizerG.step()
return errG_total
Example #15
| Source File: trainer.py From Recipe2ImageGAN with MIT License | 4 votes |
|
def train_Gnet(self, count):
self.netG.zero_grad()
errG_total = 0
flag = count % 100
batch_size = self.real_imgs[0].size(0)
criterion, mu, logvar = self.criterion, self.mu, self.logvar
real_labels = self.real_labels[:batch_size]
for i in range(self.num_Ds):
outputs = self.netsD[i](self.fake_imgs[i], mu)
errG = criterion(outputs[0], real_labels)
if len(outputs) > 1 and cfg.TRAIN.COEFF.UNCOND_LOSS > 0:
errG_patch = cfg.TRAIN.COEFF.UNCOND_LOSS *\
criterion(outputs[1], real_labels)
errG = errG + errG_patch
errG_total = errG_total + errG
if flag == 0:
summary_D = summary.scalar('G_loss%d' % i, errG.data[0])
self.summary_writer.add_summary(summary_D, count)
# Compute color consistency losses
if cfg.TRAIN.COEFF.COLOR_LOSS > 0:
if self.num_Ds > 1:
mu1, covariance1 = compute_mean_covariance(self.fake_imgs[-1])
mu2, covariance2 = \
compute_mean_covariance(self.fake_imgs[-2].detach())
like_mu2 = cfg.TRAIN.COEFF.COLOR_LOSS * nn.MSELoss()(mu1, mu2)
like_cov2 = cfg.TRAIN.COEFF.COLOR_LOSS * 5 * \
nn.MSELoss()(covariance1, covariance2)
errG_total = errG_total + like_mu2 + like_cov2
if flag == 0:
sum_mu = summary.scalar('G_like_mu2', like_mu2.data[0])
self.summary_writer.add_summary(sum_mu, count)
sum_cov = summary.scalar('G_like_cov2', like_cov2.data[0])
self.summary_writer.add_summary(sum_cov, count)
if self.num_Ds > 2:
mu1, covariance1 = compute_mean_covariance(self.fake_imgs[-2])
mu2, covariance2 = \
compute_mean_covariance(self.fake_imgs[-3].detach())
like_mu1 = cfg.TRAIN.COEFF.COLOR_LOSS * nn.MSELoss()(mu1, mu2)
like_cov1 = cfg.TRAIN.COEFF.COLOR_LOSS * 5 * \
nn.MSELoss()(covariance1, covariance2)
errG_total = errG_total + like_mu1 + like_cov1
if flag == 0:
sum_mu = summary.scalar('G_like_mu1', like_mu1.data[0])
self.summary_writer.add_summary(sum_mu, count)
sum_cov = summary.scalar('G_like_cov1', like_cov1.data[0])
self.summary_writer.add_summary(sum_cov, count)
kl_loss = KL_loss(mu, logvar) * cfg.TRAIN.COEFF.KL
errG_total = errG_total + kl_loss
errG_total.backward()
self.optimizerG.step()
return kl_loss, errG_total
Example #16
| Source File: trainer.py From Recipe2ImageGAN with MIT License | 4 votes |
|
def train_Dnet(self, idx, count):
flag = count % 100
batch_size = self.real_imgs[0].size(0)
criterion, mu = self.criterion, self.mu
netD, optD = self.netsD[idx], self.optimizersD[idx]
real_imgs = self.real_imgs[idx]
wrong_imgs = self.wrong_imgs[idx]
fake_imgs = self.fake_imgs[idx]
#
netD.zero_grad()
# Forward
real_labels = self.real_labels[:batch_size]
fake_labels = self.fake_labels[:batch_size]
# for real
real_logits = netD(real_imgs, mu.detach())
wrong_logits = netD(wrong_imgs, mu.detach())
fake_logits = netD(fake_imgs.detach(), mu.detach())
#
errD_real = criterion(real_logits[0], real_labels)
errD_wrong = criterion(wrong_logits[0], fake_labels)
errD_fake = criterion(fake_logits[0], fake_labels)
if len(real_logits) > 1 and cfg.TRAIN.COEFF.UNCOND_LOSS > 0:
errD_real_uncond = cfg.TRAIN.COEFF.UNCOND_LOSS * \
criterion(real_logits[1], real_labels)
errD_wrong_uncond = cfg.TRAIN.COEFF.UNCOND_LOSS * \
criterion(wrong_logits[1], real_labels)
errD_fake_uncond = cfg.TRAIN.COEFF.UNCOND_LOSS * \
criterion(fake_logits[1], fake_labels)
#
errD_real = errD_real + errD_real_uncond
errD_wrong = errD_wrong + errD_wrong_uncond
errD_fake = errD_fake + errD_fake_uncond
#
errD = errD_real + errD_wrong + errD_fake
else:
errD = errD_real + 0.5 * (errD_wrong + errD_fake)
# backward
errD.backward()
# update parameters
optD.step()
# log
if flag == 0:
summary_D = summary.scalar('D_loss%d' % idx, errD.data[0])
self.summary_writer.add_summary(summary_D, count)
return errD
Example #17
| Source File: blah.py From Recipe2ImageGAN with MIT License | 4 votes |
|
def train_Gnet(self, count):
self.netG.zero_grad()
errG_total = 0
flag = count % 100
batch_size = self.real_imgs[0].size(0)
criterion, mu, logvar = self.criterion, self.mu, self.logvar
real_labels = self.real_labels[:batch_size]
for i in range(self.num_Ds):
outputs = self.netsD[i](self.fake_imgs[i], mu)
errG = criterion(outputs[0], real_labels)
if len(outputs) > 1 and cfg.TRAIN.COEFF.UNCOND_LOSS > 0:
errG_patch = cfg.TRAIN.COEFF.UNCOND_LOSS *\
criterion(outputs[1], real_labels)
errG = errG + errG_patch
errG_total = errG_total + errG
if flag == 0:
summary_D = summary.scalar('G_loss%d' % i, errG.data[0])
self.summary_writer.add_summary(summary_D, count)
# Compute color consistency losses
if cfg.TRAIN.COEFF.COLOR_LOSS > 0:
if self.num_Ds > 1:
mu1, covariance1 = compute_mean_covariance(self.fake_imgs[-1])
mu2, covariance2 = \
compute_mean_covariance(self.fake_imgs[-2].detach())
like_mu2 = cfg.TRAIN.COEFF.COLOR_LOSS * nn.MSELoss()(mu1, mu2)
like_cov2 = cfg.TRAIN.COEFF.COLOR_LOSS * 5 * \
nn.MSELoss()(covariance1, covariance2)
errG_total = errG_total + like_mu2 + like_cov2
if flag == 0:
sum_mu = summary.scalar('G_like_mu2', like_mu2.data[0])
self.summary_writer.add_summary(sum_mu, count)
sum_cov = summary.scalar('G_like_cov2', like_cov2.data[0])
self.summary_writer.add_summary(sum_cov, count)
if self.num_Ds > 2:
mu1, covariance1 = compute_mean_covariance(self.fake_imgs[-2])
mu2, covariance2 = \
compute_mean_covariance(self.fake_imgs[-3].detach())
like_mu1 = cfg.TRAIN.COEFF.COLOR_LOSS * nn.MSELoss()(mu1, mu2)
like_cov1 = cfg.TRAIN.COEFF.COLOR_LOSS * 5 * \
nn.MSELoss()(covariance1, covariance2)
errG_total = errG_total + like_mu1 + like_cov1
if flag == 0:
sum_mu = summary.scalar('G_like_mu1', like_mu1.data[0])
self.summary_writer.add_summary(sum_mu, count)
sum_cov = summary.scalar('G_like_cov1', like_cov1.data[0])
self.summary_writer.add_summary(sum_cov, count)
kl_loss = KL_loss(mu, logvar) * cfg.TRAIN.COEFF.KL
errG_total = errG_total + kl_loss
errG_total.backward()
self.optimizerG.step()
return kl_loss, errG_total
Example #18
| Source File: blah.py From Recipe2ImageGAN with MIT License | 4 votes |
|
def train_Dnet(self, idx, count):
flag = count % 100
batch_size = self.real_imgs[0].size(0)
criterion, mu = self.criterion, self.mu
netD, optD = self.netsD[idx], self.optimizersD[idx]
real_imgs = self.real_imgs[idx]
wrong_imgs = self.wrong_imgs[idx]
fake_imgs = self.fake_imgs[idx]
#
netD.zero_grad()
# Forward
real_labels = self.real_labels[:batch_size]
fake_labels = self.fake_labels[:batch_size]
# for real
real_logits = netD(real_imgs, mu.detach())
wrong_logits = netD(wrong_imgs, mu.detach())
fake_logits = netD(fake_imgs.detach(), mu.detach())
#
errD_real = criterion(real_logits[0], real_labels)
errD_wrong = criterion(wrong_logits[0], fake_labels)
errD_fake = criterion(fake_logits[0], fake_labels)
if len(real_logits) > 1 and cfg.TRAIN.COEFF.UNCOND_LOSS > 0:
errD_real_uncond = cfg.TRAIN.COEFF.UNCOND_LOSS * \
criterion(real_logits[1], real_labels)
errD_wrong_uncond = cfg.TRAIN.COEFF.UNCOND_LOSS * \
criterion(wrong_logits[1], real_labels)
errD_fake_uncond = cfg.TRAIN.COEFF.UNCOND_LOSS * \
criterion(fake_logits[1], fake_labels)
#
errD_real = errD_real + errD_real_uncond
errD_wrong = errD_wrong + errD_wrong_uncond
errD_fake = errD_fake + errD_fake_uncond
#
errD = errD_real + errD_wrong + errD_fake
else:
errD = errD_real + 0.5 * (errD_wrong + errD_fake)
# backward
errD.backward()
# update parameters
optD.step()
# log
if flag == 0:
summary_D = summary.scalar('D_loss%d' % idx, errD.data[0])
self.summary_writer.add_summary(summary_D, count)
return errD
Example #19
| Source File: blah.py From Recipe2ImageGAN with MIT License | 4 votes |
|
def train_Gnet(self, count):
self.netG.zero_grad()
errG_total = 0
flag = count % 100
batch_size = self.real_imgs[0].size(0)
criterion = self.criterion
real_labels = self.real_labels[:batch_size]
for i in range(self.num_Ds):
netD = self.netsD[i]
outputs = netD(self.fake_imgs[i])
errG = criterion(outputs[0], real_labels)
# errG = self.stage_coeff[i] * errG
errG_total = errG_total + errG
if flag == 0:
summary_G = summary.scalar('G_loss%d' % i, errG.data[0])
self.summary_writer.add_summary(summary_G, count)
# Compute color preserve losses
if cfg.TRAIN.COEFF.COLOR_LOSS > 0:
if self.num_Ds > 1:
mu1, covariance1 = compute_mean_covariance(self.fake_imgs[-1])
mu2, covariance2 = \
compute_mean_covariance(self.fake_imgs[-2].detach())
like_mu2 = cfg.TRAIN.COEFF.COLOR_LOSS * nn.MSELoss()(mu1, mu2)
like_cov2 = cfg.TRAIN.COEFF.COLOR_LOSS * 5 * \
nn.MSELoss()(covariance1, covariance2)
errG_total = errG_total + like_mu2 + like_cov2
if self.num_Ds > 2:
mu1, covariance1 = compute_mean_covariance(self.fake_imgs[-2])
mu2, covariance2 = \
compute_mean_covariance(self.fake_imgs[-3].detach())
like_mu1 = cfg.TRAIN.COEFF.COLOR_LOSS * nn.MSELoss()(mu1, mu2)
like_cov1 = cfg.TRAIN.COEFF.COLOR_LOSS * 5 * \
nn.MSELoss()(covariance1, covariance2)
errG_total = errG_total + like_mu1 + like_cov1
if flag == 0:
sum_mu = summary.scalar('G_like_mu2', like_mu2.data[0])
self.summary_writer.add_summary(sum_mu, count)
sum_cov = summary.scalar('G_like_cov2', like_cov2.data[0])
self.summary_writer.add_summary(sum_cov, count)
if self.num_Ds > 2:
sum_mu = summary.scalar('G_like_mu1', like_mu1.data[0])
self.summary_writer.add_summary(sum_mu, count)
sum_cov = summary.scalar('G_like_cov1', like_cov1.data[0])
self.summary_writer.add_summary(sum_cov, count)
errG_total.backward()
self.optimizerG.step()
return errG_total
