returnn.tf.compat
#
This module provides some helpers to support earlier versions of TF 1 and also TF 2. In all cases, we use graph-mode execution, i.e. we will disable eager-mode.
For an initial draft about TF 2 support, see here: https://github.com/rwth-i6/returnn/issues/283
Also see this for a similar (but different) approach: https://github.com/tensorflow/lingvo/blob/master/lingvo/compat.py