Closed purvanshi closed 6 years ago
Are you using your own data or the datasets from CMU Multimodal SDK? Have you examined if there're NaN values in the input? For the past version of SDK I wrote code for checking and removing inf and nan values manually, it might be the case that in your data or the new version of SDK's data there're also inf or nan values in the input. In my previous experience, this happens in particular with COVAREP acoustic features.
yes the problem is with acoustic features. The input is nan at several points. Thanks for the help.
The layer1 weights of the audio subnetwork start giving NAN values. Particularly talking about the output of
self.linear_1(dropped)
I tried adjusting the learning rate but it doesn't help.