Drop in performance when converting from Keras to TFLite (float 32)


I was trying to convert my model to TFLite and it is in float 32. I see a small drop in metrics. So my question is really straightforward - is this drop expected when converting to TFLite?


Yes, a small drop in performance metrics is expected when converting a TensorFlow model to TensorFlow Lite (TFLite) format, even in float32. This can be due to optimizations for mobile devices, slight numerical differences, and variations in operator support and compatibility.

Thanks for that answer.
I got an other question - so when I am converting my model, I am using BUILTIN_REF instead of AUTO OpsResolver. I know this is used for testing and debugging but could this be a problem when converting to TFLite (on metrics, etc…) so maybe there are some difference between AUTO and BUILTIN_REF that could lead to more drops?