Updater

class Updater(module: Updatable, *params: str, reduction: Callable[[Tensor, int], Tensor] | None = None, **kwargs)

Bases: Module

Managed accumulated updates for module parameters.

The added parameters are all set as properties which return an Accumulator corresponding to that parameter. Care must be taken to avoid naming collisions, although the number of attributes in Updater not in Module are small. See the methods _getacc_(), _setacc_(), and _delacc_() for more information.

When a reduction is not specified, the default from Accumulator.reduction is used.

Parameters:

• module (Updatable) – module with updatable parameters.

• *params (str) – parameters to set as trainable.

• reduction (Callable[[torch.Tensor, int], torch.Tensor] | None, optional) – function for reducing updates. Defaults to None.

Caution

An Updater only weakly references its parent module, if its parent is deleted this updater will be made invalid.

Note

The initializer creates an object of a dynamically created type with a base type of Updater.

static _delacc_(self: Updater, attr: str) → None

Clears the accumulator state for a given attribute.

As a property, this is equivalent to using del on the Accumulator.pos and Accumulator.neg properties directly, which itself resets them back to their empty states.

Parameters:

• self (Updater) – updater, self via the associated property.

• attr (str) – parameter name to target.

static _getacc_(self: Updater, attr: str) → Accumulator

Gets the accumulator for a given attribute.

Parameters:

• self (Updater) – updater, self via the associated property.

• attr (str) – parameter name to target.

Returns:

associated accumulator for the given parameter.

Return type:

Accumulator

static _setacc_(self: Updater, value: tuple[Tensor | None, Tensor | None] | Tensor | None, attr: str) → None

Updates the accumulator values for a given attribute.

As a property, setting with a 2-tuple assumes the first term is the positive portion of the update and the second term is the negative portion. If instead a tensor is given, it assumes this update is only the positive portion. Any None values are ignored. The following blocks shows equivalent statements.

updater.attr = pos_update, neg_update
updater.attr.pos, updater.attr.neg = pos_update, neg_update

updater.attr = pos_update
updater.attr.pos = pos_update

Parameters:

• self (Updater) – updater, self via the associated property.

• value (tuple[torch.Tensor | None, torch.Tensor | None] | torch.Tensor | None) – value of the update to assign.

• attr (str) – parameter name to target.

Important

The negative portions of updates should still be positively valued as they will be subtracted from the positive portion.

clear(**kwargs) → None

Clears all of the accumulators’ states.

forward(*params: str, **kwargs) → None

Applies accumulated updates.

Parameters:

*params (str) – parameters to update, all parameters when None are specified.

property names: tuple[str, ...]

Names of updatable attributes.

Returns:

names of updatable parameters.

Return type:

tuple[str, …]

property parent: Module | None

Parent module, if valid.

Returns:

parent module if the reference to it still exists.

Return type:

Module | None