Support for tensor2tensor ranged_hparams

[cwbeitel]

Tensor2tensor is an increasingly popular ML research framework built on top of tensorflow that has built-in support for ranged hparams, e.g.

@registry.register_ranged_hparams
def rlmb_grid(rhp):
  """Grid over games and frames, and 5 runs each for variance."""
  rhp.set_categorical("loop.game", ["breakout", "pong", "freeway"])
  base = 100000
  medium = base // 2
  small = medium // 2
  rhp.set_discrete("loop.num_real_env_frames", [base, medium, small])

  # Dummy parameter to get 5 runs for each configuration
  rhp.set_discrete("model.moe_loss_coef", list(range(5)))

also here.

In https://github.com/kubeflow/katib/issues/240 it was proposed to add a python interface to launching katib jobs which would include providing a means of specifying hparam ranges (one rough method was proposed). For those using katib to tune tensor2tensor models (not everyone), the user experience for this could stay closer to that of using tensor2tensor normally (outside of tuning) if katib supported ranged_hparams as above as a way of specifying the space over which katib should tune.

Related to https://github.com/kubeflow/examples/pull/322

/cc @jlewi @texasmichelle

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[andreyvelich]

/lifecycle frozen