Calls obj with args and kwargs and return its result if obj is callable, otherwise returns obj.

Method for getting outergradient and outer parameters as required by apply_gradient methods from tensorflow optimizers. :par

reuses defined convolutional networks with new input and update the output results :param new_input: new input with same shape as the

reuses defined convolutional networks with new input and update the output results :param new_input: new input with same shape as the

reuses defined convolutional networks with new input and update the output results :param new_input: new input with same shape as the

:param x: input for the i-th block :param i: i-th block :return: the output of the j-th conv block in the i-th residual block

:param dataset: :return: just for testing

Get's the `activation` :param item: :return:

:param metadataset: instance of data set :param n_batches: number of batches :param batch_size: size of batch :param

Creates an object that creates model parameters and defines the network structure. :param _input: the input shape for defined network

:param _input: original input :param dim_output: dimension of output :param name: scope of meta-learner :param outer_

:param _input: original input :param dim_output: dimension of output :param name: scope of meta-learner :param outer_

:param _input: original input :param dim_output: dimension of output :param name: scope of meta-learner :param outer_

:param _input: original input for the FeedForward network, which is also output of the convolutional neural networks :param

BMLHOptimizer is the main class in `boml`, which takes responsibility for the whole process of model construnction and back propagati

Utility method to initialize truncated reverse HG (not necessarily online), :param truncate_iter: Maximum number of iterations that wil

Utility method to initialize truncated reverse HG (not necessarily online), :param name: a name for the operations and variables that w

Define data formats and manage attributes of the generated tasks. :param datasets: instance of Meta Dataset :param dtype: def

Class for managing windowed input data (like TIMIT). :param data: Numpy matrix. Each row should be an example data :param ro

Class for stochastic sampling of data points. It is most useful for feeding examples for the the training ops of `ReverseHG` or `Forw

Construct a new gradient descent optimizer. Args: learning_rate: A Tensor or a floating point value. The learning

r"""Construct a new Adam optimizer. Initialization: $$m_0 := 0 \text{(Initialize initial 1st moment vector)}$$ $$v_0 := 0 \t

Construct a new Momentum optimizer. Args: learning_rate: A `Tensor` or a floating point value. The learning rate.

:param update_op: the operation to perform gradient-based optimization :param dynamics: the iterative formats of dynamical system

:param update_op: the tf operation to perform updates :param dynamics: the iterative formats of dynamical system :param objec

:param update_op: the tf operation to perform updates :param dynamics: the iterative formats of dynamical system :param objec

_forward() uses defined convolutional neural networks with initial input :return:

_forward() uses defined convolutional neural networks with initial input :return:

_forward() uses defined convolutional neural networks with initial input :return:

_forward() uses defined convolutional neural networks with initial input :return:

_forward() uses defined convolutional neural networks with initial input :return:

_forward() uses defined convolutional neural networks with initial input :return:

iteration of minimization of UL objective(s), assuming the hyper-gradients are already computed.

:param step: unused, just for making it compatible with `HG` and `Saver` :return: the feed dictionary

Runs the inner optimization dynamics for T iterations in the meanwhile. :param inner_objective_feed_dicts: Optional feed dic

Apply gradients to variables. This is the second part of `minimize()`. It returns an `Operation` that applies gradients. Arg

Apply gradients to variables. This is the second part of `minimize()`. It returns an `Operation` that applies gradients. Arg

Descent step, as returned by `tf.train.Optimizer.apply_gradients`. :return:

Makes sure `obj` is a list or otherwise converts it to a list with a single element.

Makes sure `obj` is a list or otherwise converts it to a list with a single element. :param obj: :return: A `list`

Make sure that `obj` is a tuple or a list and eventually converts it into a list with a single element :param obj: :return: A `tuple` or

Function that adds to the computational graph all the operations needend for computing the hypergradients in a "dynamic" way, without

Function overridden by specific methods. :param boml_inner_grad: inner_grad object resulting from the inner objective optimization.

Function that adds to the computational graph all the operations needed for computing the outer gradients with the dynamical system.

Function that adds to the computational graph all the operations needend for computing the hypergradients in a "dynamic" way, without

:param boml_opt: instance of modified optimizers in the `optimizer` module :param loss_inner: Lower-Level objectives :param l

:param boml_opt: instance of modified optimizers in the `optimizer` module :param loss_inner: Lower-Level objectives :param l

Perform, conv, batch norm, nonlinearity, and max pool :param boml_net: instance of BOMLNet :param convweight: parameter of convolutional filt

Perform, conv, batch norm, nonlinearity, and max pool :param boml_net: instance of BOMLNet :param cweight: parameter of convolutional filter