Open annzheng opened 5 years ago
Hi, I found in function _sigmoid(), you clamp the value between (1e-4, 1-1e-4) after sigmoid, why need to do this? Here is the code: def _sigmoid(x): x = torch.clamp(x.sigmoid_(), min=1e-4, max=1-1e-4) return x
def _sigmoid(x):
x = torch.clamp(x.sigmoid_(), min=1e-4, max=1-1e-4)
return x
We did it to make the training more stable.
Hi @heilaw!
How does it make the training more stable?
Hi, I found in function _sigmoid(), you clamp the value between (1e-4, 1-1e-4) after sigmoid, why need to do this? Here is the code:
def _sigmoid(x):
x = torch.clamp(x.sigmoid_(), min=1e-4, max=1-1e-4)
return x