def log_rank_loss(self, y_pos, y_neg, temp=0):
M = y_pos.size(0)
N = y_neg.size(0)
y_pos = self.gamma - y_pos
y_neg = self.gamma - y_neg
C = int(N / M)
y_neg = y_neg.view(C, -1).transpose(0, 1)
p = F.softmax(temp * y_neg)
loss_pos = torch.sum(F.softplus(-1 * y_pos))
loss_neg = torch.sum(p * F.softplus(y_neg))
loss = (loss_pos + loss_neg) / 2 / M
if self.gpu:
loss = loss.cuda()
return loss
def g_nonsaturating_loss(self, fake_pred):
loss = F.softplus(-fake_pred).mean()
return loss
def d_logistic_loss(self, real_pred, fake_pred):
real_loss = F.softplus(-real_pred)
fake_loss = F.softplus(fake_pred)
return (real_loss + fake_loss).mean()
def forward(x):
return x * F.softplus(x).tanh()
def backward(ctx, grad_output):
x = ctx.saved_tensors[0]
sx = torch.sigmoid(x)
fx = F.softplus(x).tanh()
return grad_output * (fx + x * sx * (1 - fx * fx))
def forward(ctx, x):
ctx.save_for_backward(x)
return x.mul(torch.tanh(F.softplus(x))) # x * tanh(ln(1 + exp(x)))
def forward(self, x):
return x * (torch.tanh(F.softplus(x)))
def latent_param(self): return self.latent_mean.mul(1), F.softplus(self.latent_logvar).div(math.log(2)).mul(1)
