What Are _get_hyper And _set_hyper In Tensorflow Optimizers?

I see it in __init__ of e.g. Adam optimizer: self._set_hyper('beta_1', beta_1). There are also _get_hyper and _serialize_hyperparameter throughout the code. I don't see these in Ke

Solution 1:

They enable setting and getting Python literals (int, str, etc), callables, and tensors. Usage is for convenience and consistency: anything set via _set_hyper can be retrieved via _get_hyper, avoiding repeating boilerplate code. I've implemented Keras AdamW in all major TF & Keras versions, and will use it as reference.

• t_cur is a tf.Variable. Each time we "set" it, we must invoke K.set_value; if we do self.t_cur=5, this will destroy tf.Variable and wreck optimizer functionality. If instead we used model.optimizer._set_hyper('t_cur', 5), it'd set it appropriately - but this requires for it to have been defined via set_hyperpreviously.

• Both _get_hyper & _set_hyper enable programmatic treatment of attributes - e.g., we can make a for-loop with a list of attribute names to get or set using just _get_hyper and _set_hyper, whereas otherwise we'd need to code conditionals and typechecks. Also, _get_hyper(name) requires that name was previously set via set_hyper.

• _get_hyper enables typecasting via dtype=. Ex: beta_1_t in default Adam is cast to same numeric type as var (e.g. layer weight), which is required for some ops. Again a convenience, as we could typecast manually (math_ops.cast).

• _set_hyper enables the use of _serialize_hyperparameter, which retrieves the Python values (int, float, etc) of callables, tensors, or already-Python values. Name stems from the need to convert tensors and callables to Pythonics for e.g. pickling or json-serializing - but can be used as convenience for seeing tensor values in Graph execution.

• Lastly; everything instantiated via _set_hyper gets assigned to optimizer._hyper dictionary, which is then iterated over in _create_hypers. The else in the loop casts all Python numerics to tensors - so _set_hyper will not create int, float, etc attributes. Worth noting is the aggregation= kwarg, whose documentation reads: "Indicates how a distributed variable will be aggregated". This is the part a bit more than "for convenience" (lots of code to replicate).

  • _set_hyper has a limitation: does not allow instantiating dtype. If add_weight approach in _create_hypers is desired with dtype, then it should be called directly.

When to use vs. not use: use if the attribute is used by the optimizer via TensorFlow ops - i.e. if it needs to be a tf.Variable. For example, epsilon is set regularly, as it's never needed as a tensor variable.