ddf.ant_link package

Submodules

ddf.ant_link.ant_link module

class ddf.ant_link.ant_link.AntLink(id: ddf.ddf.Id)

Bases: ABC

backpropagate(node: Node, gradient: ndarray, losstype: LossChannel) → None

generate_gradient_messages(node: Node, gradient_storage: GradientStorage) → list[GradientMessage]

id: Id

optimize_step(gradient_storage: GradientStorage, object_id: ObjectId, exchange_vector: ExchangeVector) → None | ExchangeVector

Return updated values for state_vec and state_cov.

class ddf.ant_link.ant_link.GradientSubscription(loss_channel: ddf.ant_link.loss_channel.LossChannel, solver: collections.abc.Callable[[numpy.ndarray, ddf.information.ExchangeVector, dict], tuple[ddf.information.ExchangeVector, dict]], solver_state: dict)

Bases: object

loss_channel: LossChannel

solver: Callable[[ndarray, ExchangeVector, dict], tuple[ExchangeVector, dict]]

solver_state: dict

ddf.ant_link.ant_link.Rprop_step(grad: ndarray, exchange_vector: ExchangeVector, delta_0: ndarray, delta_min: ndarray, delta_max: ndarray, solver_dict: dict, lb: ndarray = array([], dtype=float64), ub: ndarray = array([], dtype=float64)) → tuple[ExchangeVector, dict]

Calculate a gradient descent step.

Calculate a gradient descent step in the device parameters while controlling the learning rate with the iRProp-algorithm.

Respects the provided parameter bounds.

See C. Igel and Michael H¨usken. Improving the rprop learning algorithm. In Proceedings of the Second International Symposium on Neural Computation, NC2000, 2000.

Parameters:

• grad (1D nd array) – loss gradient

• exchange_vector (ExchangeVector) – current parameter vector

• delta_0 (1D nd array) – initial step size

• delta_min (1D nd array) – minimum step size

• delta_max (1D nd array) – maximum step size

• solver_dict (dict) – dict for transfering temporary variables form iteration to iteration

• lb (1D ndarray) – parameterlower bound vector

• ub (1D ndarray) – parameter upper bound vector

Returns:

new parameter vector after update dict: the solver_dict to pass in the next iteration

Return type:

1D ndarray

ddf.ant_link.external module

class ddf.ant_link.external.AntLinkExternal(id: ~ddf.ddf.Id, update_counter: int, gradient_threshold: float = 1e-06, max_gradients: dict[~ddf.ant_link.loss_channel.LossChannel, ~numpy.ndarray] = <factory>)

Bases: AntLink

Refer to a different ObjectId on a different Node.

generate_gradient_messages(node: Node, gradient_storage: GradientStorage) → list[GradientMessage]

generate_info_message(node: Node, state: ExchangeVector) → InfoMessage

Generate an info message.

gradient_threshold: float = 1e-06

max_gradients: dict[LossChannel, ndarray]

update_counter: int

ddf.ant_link.internal module

class ddf.ant_link.internal.AntLinkInternalObjectID(id: Id)

Bases: AntLink

Refer to a different ObjectId on the same Node.

backpropagate(node: Node, gradient: ndarray, losstype: LossChannel) → None

Gradient w.r.t. the exchange vector.

ddf.ant_link.local module

class ddf.ant_link.local.AntLinkLocal(id: Id, gradient_subscriptions: list[GradientSubscription])

Bases: AntLink

Refer to the same ObjectId on the same Node.

gradient_subscriptions: list[GradientSubscription]

optimize_step(gradient_storage: GradientStorage, object_id: ObjectId, exchange_vector: ExchangeVector) → None | ExchangeVector

Return updated values for state_vec and state_cov.

ddf.ant_link.loss_channel module

class ddf.ant_link.loss_channel.FuturePastProbe(*values)

Bases: Enum

FUTURE = 0

PAST = 1

PROBE = 2

class ddf.ant_link.loss_channel.LossChannel(mode: ddf.ant_link.loss_channel.FuturePastProbe, lossref: uuid.UUID | None, name: str)

Bases: object

classmethod from_string(loss_channel: str) → LossChannel

lossref: UUID | None

mode: FuturePastProbe

name: str

to_string() → str

Module contents