tf.ConfigProto usage on TF 2.0

[samsamoa]

On TensorFlow 1.X, there are various important parameters set by passing tf.ConfigProto to tf.Session(config=...) or tf.enable_eager_execution(config=...). For example, to use NCCL, it is useful to set the visible GPUs for a session with config.gpu_options.visible_device_list.

My understanding is that TensorFlow 2.0 no longer has a way to set this configuration — both tf.Session and tf.enable_eager_execution are gone. Is there an alternate way to set this config?

Related StackOverflow question
CC @girving @allenlavoie

[dynamicwebpaige]

Potentially related: #25138.

[martinwicke]

Gaurav is working on a proper replacement, it's not ready yet, for now please use compat.v1.

[girving]

We don't know how to use compat.v1 to do this. If there is a way, would it be possible to answer the Stackoverflow question so that it would be possible for us to try TF 2? The implication from @allenlavoie was that no workaround existed.

[dynamicwebpaige]

@girving - If you use the upgrade script on your original TensorFlow model, tf.Session(config=...) should change to tf.compat.v1.Session(config=...) and work as expected.

[martinwicke]

(and you should be able to use tf.compat.v1.ConfigProto to create the config parameter)

[samsamoa]

If I want to experiment with writing "true TF 2.0 code", will tf.compat.v1.enable_eager_execution(config=...) have the desired effect? In other words, should I expect that we'll end up with a drop-in replacement for the above command?