Add Hyperopt example for BERT classifier

[AvinashBukkittu]

This PR

• Adds an example to showcase hyperparameter tuning with Executor

• Fixes a memory leakage issue in Executor due to atexit (Now, opening and closing of files are done explicitly in train and test methods)

[ZhitingHu]

So now the OOM issue is fixed?

[AvinashBukkittu]

So now the OOM issue is fixed?

Yes, the issue was caused because _close_files method (due to library atexit) was holding a reference to Executor instance due to which Adam was not cleared from memory. We removed atexit dependency.

[ZhitingHu]

Looks good on my side.

Do we still have a plan of adding relevant Executor hyper tuning modules?

[AvinashBukkittu]

Looks good on my side.

Do we still have a plan of adding relevant Executor hyper tuning modules?

I discussed with @huzecong on keras-tuner like interfaces for Executor. They do not provide any additional functionalities but may provide some good utilities for the user. I think this is good and we can add these features in the future if we get requests from the users.

[ZhitingHu]

Looks good on my side.
Do we still have a plan of adding relevant Executor hyper tuning modules?

I discussed with @huzecong on keras-tuner like interfaces for Executor. They do not provide any additional functionalities but may provide some good utilities for the user. I think this is good and we can add these features in the future if we get requests from the users.

Let's make an issue to keep track of this

[AvinashBukkittu]

Looks good on my side.
Do we still have a plan of adding relevant Executor hyper tuning modules?

I discussed with @huzecong on keras-tuner like interfaces for Executor. They do not provide any additional functionalities but may provide some good utilities for the user. I think this is good and we can add these features in the future if we get requests from the users.

Let's make an issue to keep track of this

#191