Merge branch 'fix-bug-with-mean-loss-disc' into 'dev'

Fix bug with mean loss disc

See merge request tobifinn/torch-assimilate!55

pytassim/toolbox/discriminator/standard.py

@@ -59,7 +59,7 @@ class StandardDisc(object):
  """
  def __init__(self, net,):
  self.net = net
- self.loss_func = torch.nn.BCEWithLogitsLoss()
+ self.loss_func = torch.nn.BCEWithLogitsLoss(reduction='mean')
  self.optimizer = None
  self.grad_optim = True
 
@@ -150,6 +150,7 @@ def disc_loss(self, in_data, labels):
  loss has the same tensor type as given `in_data`.
  """
  loss = self.loss_func(in_data, labels)
+ #loss = unscaled_loss * in_data.nelement() / in_data.shape[0]
  return loss
 
  def forward(self, *args, **kwargs):
@@ -179,11 +180,11 @@ def _get_train_losses(self, real_data, fake_data, *args, **kwargs):
  batch_size = real_data.size()[0]
 
  real_critic = self.forward(real_data, *args, **kwargs)
- real_labels = self.get_targets(batch_size, 1.0, real_data)
+ real_labels = self.get_targets(batch_size, 0.0, real_data)
  real_loss = self.disc_loss(real_critic, real_labels)
 
  fake_critic = self.forward(fake_data, *args, **kwargs)
- fake_labels = self.get_targets(batch_size, 0.0, real_data)
+ fake_labels = self.get_targets(batch_size, 1.0, real_data)
  fake_loss = self.disc_loss(fake_critic, fake_labels)
 
  total_loss = real_loss + fake_loss
@@ -329,8 +330,7 @@ def gen_loss(self, fake_data, *args, **kwargs):
  batch_size = fake_data.size()[0]
 
  fake_critic = self.forward(fake_data, *args, **kwargs)
- real_labels = self.get_targets(batch_size, 1.0, fake_data)
- gen_loss = self.disc_loss(fake_critic, real_labels)
+ gen_loss = torch.mean(fake_critic)
  return gen_loss
 
  def recon_loss(self, recon_obs, *args, **kwargs):

pytassim/toolbox/discriminator/zero_penalty.py

@@ -66,12 +66,12 @@ def _get_train_losses(self, real_data, fake_data, *args, **kwargs):
 
  real_data.requires_grad_()
  real_critic = self.forward(real_data, *args, **kwargs)
- real_labels = self.get_targets(batch_size, 1.0, real_data)
+ real_labels = self.get_targets(batch_size, 0.0, real_data)
  real_loss = self.disc_loss(real_critic, real_labels)
 
  fake_data.requires_grad_()
  fake_critic = self.forward(fake_data, *args, **kwargs)
- fake_labels = self.get_targets(batch_size, 0.0, real_data)
+ fake_labels = self.get_targets(batch_size, 1.0, real_data)
  fake_loss = self.disc_loss(fake_critic, fake_labels)
 
  total_loss = real_loss + fake_loss