drytorch.lib.aggregators

Module containing generic accumulator-based aggregators.

Classes

`AbstractAccumulator`()	Stateful aggregation container.
`AbstractAggregator`(**kwargs)	Aggregate named values using accumulator objects.
`Averager`(**kwargs)	Aggregator computing mean over floats.
`MeanAccumulator`(total, count)	Accumulator computing arithmetic mean for floats.
`TorchAverager`(**kwargs)	Aggregator computing mean over tensors with distributed support.
`TorchMeanAccumulator`(total, count)	Accumulator computing arithmetic mean for tensors.

class AbstractAccumulator[source]

Bases: Generic[_T, _R], ABC

Stateful aggregation container.

abstractmethod classmethod from_value(value: _T) → Self[source]

Create accumulator from raw value.

Parameters:: value (_T)
Return type:: Self

abstractmethod merge(other: Self) → None[source]

Merge another accumulator into this one.

Parameters:: other (Self)
Return type:: None

abstractmethod reduce() → _R[source]

Return reduced value.

Return type:: _R

abstractmethod sync() → None[source]

Synchronize state across distributed processes.

Return type:: None

class AbstractAggregator(**kwargs: _T)[source]

Bases: Generic[_T, _R]

Aggregate named values using accumulator objects.

Initialize.

Parameters:: kwargs (_T) – named values to aggregate.

__add__(other: Self | Mapping[str, _T]) → Self[source]

Return new aggregator containing merged data.

Parameters:: other (Self | Mapping[str, _T])
Return type:: Self

__bool__() → bool[source]

Return True if any values are stored.

Return type:: bool

__iadd__(other: Self | Mapping[str, _T]) → Self[source]

Merge another aggregator or mapping into this one.

Parameters:: other (Self | Mapping[str, _T])
Return type:: Self

__repr__() → str[source]

Return representation of the aggregator.

Return type:: str

clear() → None[source]

Remove all accumulated data.

Return type:: None

keys() → list[str][source]

Return stored metric names.

Return type:: list[str]

reduce() → dict[str, _R][source]

Return reduced values for all metrics.

Return type:: dict[str, _R]

all_reduce() → dict[str, _R][source]

Synchronize accumulators across processes and reduce.

Return type:: dict[str, _R]

class MeanAccumulator(total: float, count: int)[source]

Bases: AbstractAccumulator[float, float]

Accumulator computing arithmetic mean for floats.

Parameters:

total (float)
count (int)

total

sum of all values.

Type:: float

count

number of values.

Type:: int

classmethod from_value(value: float) → MeanAccumulator[source]

Create accumulator from raw value.

Parameters:: value (float)
Return type:: MeanAccumulator

merge(other: Self) → None[source]

Merge another accumulator into this one.

Parameters:: other (Self)
Return type:: None

reduce() → float[source]

Return reduced value.

Return type:: float

sync() → None[source]

Synchronize state across distributed processes.

Return type:: None

class TorchMeanAccumulator(total: Tensor, count: int)[source]

Bases: AbstractAccumulator[Tensor, Tensor]

Accumulator computing arithmetic mean for tensors.

Parameters:

total (Tensor)
count (int)

total

sum of all values.

Type:: torch.Tensor

count

number of values.

Type:: int

classmethod from_value(value: Tensor) → TorchMeanAccumulator[source]

Create accumulator from raw value.

Parameters:: value (Tensor)
Return type:: TorchMeanAccumulator

merge(other: Self) → None[source]

Merge another accumulator into this one.

Parameters:: other (Self)
Return type:: None

reduce() → Tensor[source]

Return reduced value.

Return type:: Tensor

sync() → None[source]

Synchronize state across distributed processes.

Return type:: None

class Averager(**kwargs: _T)[source]

Bases: AbstractAggregator[float, float]

Aggregator computing mean over floats.

Initialize.

Parameters:: kwargs (_T) – named values to aggregate.

accumulator_cls: alias of MeanAccumulator

class TorchAverager(**kwargs: _T)[source]

Bases: AbstractAggregator[Tensor, Tensor]

Aggregator computing mean over tensors with distributed support.

Initialize.

Parameters:: kwargs (_T) – named values to aggregate.

accumulator_cls: alias of TorchMeanAccumulator