Jake Schmidt

Is anyone else experiencing this issue?

Mixed precision training is common / standard these days — without mixed precision support for Adasum, it feels like Adasum is a nonstarter.

@rmothukuru Yes I was aware of `generate_statistics_from_tfrecord` (and I'll update the original post). Why do I need to write serving data to tfrecords just to check for anomalies?

In the limit, yes. Serving examples could come in one at a time, or in batches. Thus the idea to have a tensor-based example validator, as the serving environment (tensorflow...

I'm also needing to delete stuff.

Typically the vocab is computed/known in an upstream step... would it be the worst idea to update the EvalConfig proto to have a field for `vocab`?

@mdreves any thoughts about this suggestion?