Why are not used log(probabilities)?
janvainer opened this issue · comments
Jan Vainer commented
Thank you for this awesome repo! ;)
I was wondering why are not used log probabilities? Is the beam search stable even for long sequences?
Harald Scheidl commented
- for my use-case (text recognition), I usually had something around 100 time-steps, for which I did not run into numerical issues
- there already was a discussion about it, maybe the changes are already implemented in the fork, see: #13
- if you're comfortable with C++, it should no be too difficult to implement the changes
Jan Vainer commented
Thanks for response, I will look into the forked repo. :)
Jerod Weinman commented
@lordofluck FYI, I never did make any log-space changes. But as @githubharald says, the conversion process would be moderately straightforward
Jan Vainer commented
@weinman Thank you for the info, I am testing the code on long sequences now. If the decoding fails for my case, I may implement the log-space operations in the future