inconsistent values when called on the same inputs

#2
by isaacm - opened

When calling the model multiple times on the same input, outputs show large variance (see example below). I'm calling it on a 10-second segment of audio, normalized and prepared as instructed by the model card. Is this a known issue with the model? Any input is appreciated.

image.png

Owner

This should not happen. Please double check that "mean" and "std" are define correctly. Also, while the model should be in evaluation mode by default, you might have to force it using "model.eval()" after you load the model.

I was not able to replicate the issue on the same input.
Screenshot from 2024-08-19 13-56-20.png

Hello,
manually changing the model to eval with model.eval() fixed it. Thank you so much!

isaacm changed discussion status to closed

Sign up or log in to comment