I am training an IMLE model for generative modelling, where at some layer l some noise is added to the node values and then sent further. While calculating/updating the gradients, do the model need to know the actual values of noise?
The problem with IMLE is that it tries multiple random noise samples and then uses only one of these samples, so the PyTorch model might not have the current values corresponding to the noise sample chosen.
If we are ignoring the noise while computing gradients, any insight on how model handles it would be appreciated.