Find the right embedding layer for mismatched cases #4179
Conversation
|
The model test failure here is actually a bug in the test that I need to fix. It's a one-line change in the model repo. What's the process for doing this, given that we're running model tests here, and I can't add a fix to this PR? |
|
@matt-gardner just make a separate PR in the models repo. We can merge that into master and then re-run this workflow |
|
But that PR will fail tests without this change. |
|
You could temporarily update the workflow in models CI to use your patch branch by modifying the |
|
That seems pretty bad - I wouldn't want to merge such a PR into that repo. I can see modifying the allennlp commit in the models repo to demonstrate that we have two PRs, one on each repo, that both work, before either one gets merged. But I think we're going to have to live with the fact the we'll have a short window where one or the other CI is broken. I'm not sure there's a reasonable alternative. @dirkgr, do you see a way around this, either? |
|
Personally I'm okay with one repo being broken for a few minutes |
|
Me too. I'll make another PR for the model repo, to show that we have a fix for the breakage in this PR, then after this PR gets merged I can update the other PR to remove the CI hack, and merge the fix for the model repo tests. |
|
Ok, corresponding model PR: allenai/allennlp-models#38. |
I was diagnosing an error in the interpret code for a scibert NER model, and found this. Comments in the code should explain the problem and solution well enough, let me know if they don't.