Changes in version 0.3.2.9000                      

  - Fixed an error where autograph({expr}) would error when exporting
    undef active binding symbols.

  - for in an eager autograph() context gains the ability to process
    arbitrary python iterables, including
    tf.distribute.DistributedDataset.

  - for now automatically infers the variables of nested for loops as
    loop vars.

  - Improved error message when a break or next is signaled as a graph
    node but the intended loop is being evaluated eagerly.

                 Changes in version 0.3.2 (2021-09-17)                  

  - Added back compatibility with R-3.3 and R-3.4

                 Changes in version 0.3.1 (2021-08-19)                  

  - Fixed issue that prevented the package from being loaded by
    devtools::load_all()

                 Changes in version 0.3.0 (2021-07-14)                  

  - Improved handling of shape_invariants supplied to ag_while_opts. A
    named list of user variable shapes can be passed directly now,
    without requiring users to manually specify shapes of internal loop
    tracking tensors.
  - Deprecated back_prop arg in ag_while_opts()
  - Move 'tensorflow' package from 'Imports' to 'Suggests' to avoid
    circular dependency.
  - Added a NEWS.md file to track changes to the package.