Strategy for Robustness #946
tom-huntington
started this conversation in
Ideas
Replies: 1 comment
-
read the stackexchange question for the context: I suppose when calculating the variance, it's best not to first take the argmax. But you'll need to first take the argmax to evaluate the longest strictly increasing subsequence. |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
I encountered repetition looping on the last few words, and have just been looking at the whisper code and paper. And thought I would share a different idea.
I decided to plot the time stamps of each token
You can see the repetition looping starting around the 83rd token.
The best way to achieve robustness is usually using statistics. Perhaps there is some statistical way to evaluate outliers of strictly non-decreasing sequences. https://stats.stackexchange.com/q/604815/356970
Regardless any rough heuristic should work too.
Beta Was this translation helpful? Give feedback.
All reactions