Activate eval mode in forward

autograd
1

I’m calling self.eval() in the forward() of a nn.Module object, the loss becomes NaN.
Is this normal?

2

No, quite likely there is something else amiss.
Calling self.eval() in the forward is likely not the ideal way of doing things, either. But to say something about a better way, it would be good if you could share what should happen that led you to do that.

Best regards

Thomas

3

Thank you, Thomas!

I want to calculate the value in the evaluation mode, so that it could be compared with the regular forward( ) output.

I just figure out a workaround: copying the code of the evaluation mode, and use an boolean variable to control the execution. It works.