meta.py

The main function for defining the meta-optimizer computational graph and set up meta-training pipeline.

Functions:

• _nested_assign(ref, value)

  • Returns a nested collection of TensorFlow assign operations.
  • Arguments:
    • ref: Nested collection of TensorFlow variables.
    • value: Values to be assigned to the variables. Must have the same structure as ref.

• _nested_variable(init, name=None, trainable=False)

  • Returns a nested collection of TensorFlow variables.
  • Arguments:
    • init: Nested collection of TensorFlow initializers.
    • name: Variable name.
    • trainable: Make variables trainable (False by default).

• _wrap_variable_creation(func, custom_getter)

  • Provides a custom getter for all variable creations.

• _get_variables(func):

  • Calls func, returning any variables created, but ignoring its return value.

  • Arguments:

    • func: Function to be called.

  • Return:

    • A tuple (variables, constants) where the first element is a list of trainable variables and the second is the non-trainable variables.

• _make_with_custom_variables(func, variables):

  • Calls func and replaces any trainable variables.

• _make_nets(variables, config, net_assignments)::

  • Creates the optimizer networks.

  • Arguments:

    • variables: A list of variables to be optimized.
    • config: A dictionary of network configurations, each of which will be passed to networks.Factory to construct a single optimizer net.
    • net_assignments: A list of tuples where each tuple is of the form (netid, variable_names) and is used to assign variables to networks. netid must be a key in config.

  • Return:

    • A tuple (nets, keys, subsets) where nets is a dictionary of created optimizer nets such that the net with key keys[i] should be applied to the subset of variables listed in subsets[i].

Classes:

• class MetaOptimizer(object):

  • Learning to learn (meta) optimizer. Optimizer which has an internal RNN which takes as input, at each iteration, the gradient of the function being minimized and returns a step direction. This optimizer can then itself be optimized to learn optimization on a set of tasks.

  • Attributes:

    • meta_loss(self,make_loss,len_unroll,net_assignments=None,model_path = None,second_derivatives=False)

    • Arguments:

      • make_loss: Callable which returns the optimizee loss; note that this should create its ops in the default graph.
      • len_unroll: Number of steps to unroll.
      • net_assignments: variable to optimizer mapping. If not None, it should be a list of (k, names) tuples, where k is a valid key in the kwargs passed at at construction time and names is a list of variable names. second_derivatives: Use second derivatives (default is false).

    • Return:

      • loss: The tensor for the meta loss.
      • update: The variable for updating optimization trajectory.
      • reset: The variable for reset optimizee.
      • fx_final: The variable for final objective function value.
      • x_final: The variable for final decision variable value.
      • sub_constants: The variabler for optimizee constants.

    • meta_minimize(self, make_loss, len_unroll, learning_rate=0.01, **kwargs)

      • Returns an operator minimizing the meta-loss

      • Arguments:

        • make_loss: Callable which returns the optimizee loss; note that this should create its ops in the default graph.
        • len_unroll: Number of steps to unroll.
        • learning_rate: Learning rate for the Adam optimizer.
        • **kwargs: keyword arguments forwarded to meta_loss.

      • Return:

        • namedtuple containing (step, update, reset, fx, x)