I was verifying the gradients on one of my models and wanted to increase the precision from Float to Double to see how it affected the error on the calculation (recommended here). My model is using TransformerEncoder and I saw that it has Scalar type hard-coded to Float, with a note that it can be made into a generic constraint when TF-427 is resolved.
I was verifying the gradients on one of my models and wanted to increase the precision from
Float
toDouble
to see how it affected the error on the calculation (recommended here). My model is using TransformerEncoder and I saw that it hasScalar
type hard-coded toFloat
, with a note that it can be made into a generic constraint when TF-427 is resolved.I thought perhaps it was referring to https://github.com/tensorflow/swift/issues#427 or https://github.com/tensorflow/swift-models/issues#427, but they appear to be unrelated. Is this an internal issue tracker? I'm wondering what the TF-427 issue is that prevents it from being made into a generic constraint on
Scalar
.