Skip to content

Conversation

@JackTemaki
Copy link
Collaborator

This was not tested yet from my side, but I think there should be no significant issues.

@JackTemaki
Copy link
Collaborator Author

Still, I wonder if we even want to have the same logic for both, or if we use this chance to remove the heuristics.

@albertz
Copy link
Member

albertz commented Apr 25, 2023

Still, I wonder if we even want to have the same logic for both, or if we use this chance to remove the heuristics.

What heuristics do you mean? I only see a single heuristic there, which is that if min(score) == max(score) for some key, it would not take that key into account when checking for the N-best checkpoints, as it would be totally random anyway. Which is not really so problematic, I think.

Or do you mean the defaults, e.g. for what models to keep by default? So, what defaults do you suggest instead? Or no defaults at all and force the user to configure it?

@JackTemaki
Copy link
Collaborator Author

Still, I wonder if we even want to have the same logic for both, or if we use this chance to remove the heuristics.

What heuristics do you mean? I only see a single heuristic there, which is that if min(score) == max(score) for some key, it would not take that key into account when checking for the N-best checkpoints, as it would be totally random anyway. Which is not really so problematic, I think.
Or do you mean the defaults, e.g. for what models to keep by default? So, what defaults do you suggest instead? Or no defaults at all and force the user to configure it?

Yes that, but this might be not nice to separate for the backends now.

@albertz
Copy link
Member

albertz commented Jul 11, 2023

This was implemented independently now via #1360, so this here is obsolete.

@albertz albertz closed this Jul 11, 2023
@albertz albertz deleted the nick-unify-model-cleanup branch July 11, 2023 11:18
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants