Frank Seide
Frank Seide
It seems the right solution is to duplicate the word weights over their respective pieces? Get Outlook for iOS ________________________________ From: leiguo2018 Sent: Tuesday, July 16, 2019 1:30:38 PM To:...
Sorry, I don't remember any more than I wrote in the original issue submission. It's just so bizarre. Hieu, if InputFileStream defaults to reading text, then that's probably a bug...
Yes. We do have corpora that have malformed UTF-8 characters. That is fine if we generate plain-text vocabs, but we should not create malformed files as a principle. I don't...
No, this is a close on read. It cannot fail. Don't add a check. Close is a destructor, there should be no error condition (other than you are trying to...
It may help more on a GPU than on a CPU, as GPUs have more cores and are therefore more bottlenecked by memory bandwidth.
This is related to a discussion that Marcin and I just had literally 30 minutes ago regarding throwing exceptions instead of aborting. Since Marian can be used as a library,...
OK, this is on our TODO list. However, it will be a while, since we want to make such an intrusive change when we don't have lots of unmerged branches.
Why is this needed for offline decoding anyway? How about just keeping the option in training, but not automatically propagating it to the decoder, and defaulting to SIZE_MAX in decoding...
I would go by what the dominant use case is. If that is translation, there should be a warning if a limit is set (!), as unnoticed truncation would affect...