Adds tf.train.experimental.PythonState
Allows users to hook into TF's object-based checkpointing with arbitrary Python state. Doing this now for the TF Agents migration. This is a reasonable place to cutoff and expose a public API; the NumPy stuff can be copied out of contrib. Experimental for now pending Trackable being exposed, since it may make more sense to combine the Variable/Tensor saving API with the non-Tensor Python state API. PiperOrigin-RevId: 235579652
Loading
Please sign in to comment