You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
My main goal now is to run training, so I'll roll it back in my branch. If we want to keep the new name we should:
figure out why the validation on Train action fails with it: data/target-stage must be equal to one of the allowed values (see docs/taskcluster.md on how to run the train action)
update config templates in taskcluster/configs
probably update some other configs, for example the ones in test/fixtures and tascluster/test/params
Overall if this renaming is not that important, maybe it's easier to roll it back. It would also break all our old configs which is maybe not a big deal because we're not planning to use them as is but still.
Run
task config-generator -- zh en --name cjk
and then run a train action with it. It generates 0 tasks.The issue is apparently in renaming "target-stage: all" to "all-pipeline" which doesn't pass validation.
Also
fp16: true
doesn't pass validation, it should be either quoted or removed since it duplicates the default decoding config.The text was updated successfully, but these errors were encountered: