The big problem we have with tensorflow that means we're trying to move away from it the undocumented breaking changes. A model saved in 2.0 should work in any

The big problem we have with tensorflow that means we're trying to move away from it the undocumented breaking changes. A model saved in 2.0 should work in any later 2.x version. In practice, they've changed the implementation of the batch normalisation layer at least twice, causing 3 sig fig differences in a unet. They just don't seem to care about production use of tensorflow.