kaira.models.binary.soft_bit_thresholding.DynamicThresholder

Inheritance diagram for DynamicThresholder

class kaira.models.binary.soft_bit_thresholding.DynamicThresholder(decay: float = 0.9, initial_threshold: float = 0.5, input_type: InputType = InputType.PROBABILITY, adaptation_method: str = 'mean', bias: float = 0.0, min_threshold: float = 0.1, max_threshold: float = 0.9, *args: Any, **kwargs: Any)

Bases: SoftBitThresholder

Thresholder with dynamically adjusting threshold for non-stationary signals.

This thresholder adapts to changing signal conditions over time using exponential moving averages. It’s particularly useful for systems with time-varying noise or signal characteristics.

The dynamic threshold is computed as a weighted average of past input statistics and can adapt to gradual changes in the signal distribution.

Example

With decay=0.9, initial_threshold=0.5: The threshold will gradually adapt to the mean of the input signal.

Methods

__init__	Initialize the dynamic thresholder.
forward	Apply dynamic thresholding to convert soft bit values to hard decisions.
from_config	Create model instance from configuration.
from_hydra_config	Create model from Hydra DictConfig.
from_pretrained_config	Create model from Hugging Face PretrainedConfig.
reset_stats	Reset the running statistics.
to	Move the model to the specified device.

__init__(decay: float = 0.9, initial_threshold: float = 0.5, input_type: InputType = InputType.PROBABILITY, adaptation_method: str = 'mean', bias: float = 0.0, min_threshold: float = 0.1, max_threshold: float = 0.9, *args: Any, **kwargs: Any)

Initialize the dynamic thresholder.

Parameters:

• decay – Exponential decay factor (0-1) controlling adaptation speed. Higher values make adaptation slower but more stable.

• initial_threshold – Starting threshold value.

• input_type – Type of input values (‘prob’ or ‘llr’).

• adaptation_method – Method to adapt threshold (‘mean’, ‘median’, ‘percentile’).

• bias – Fixed bias to add to computed threshold.

• min_threshold – Minimum allowed threshold value.

• max_threshold – Maximum allowed threshold value.

• *args – Variable positional arguments passed to the base class.

• **kwargs – Variable keyword arguments passed to the base class.

classmethod from_config(config, **kwargs)

Create model instance from configuration.

Parameters:

• config – Configuration object (PretrainedConfig, DictConfig, or dict)

• **kwargs – Additional parameters to override config

Returns:

Model instance

classmethod from_hydra_config(config: DictConfig, **kwargs)

Create model from Hydra DictConfig.

Parameters:

• config – Hydra configuration

• **kwargs – Additional parameters

Returns:

Model instance

classmethod from_pretrained_config(config: PretrainedConfig, **kwargs)

Create model from Hugging Face PretrainedConfig.

Parameters:

• config – PretrainedConfig instance

• **kwargs – Additional parameters

Returns:

Model instance

to(device: str | device, *args, **kwargs) → SoftBitThresholder

Move the model to the specified device.

Parameters:

• device – The device to move the model to.

• *args – Additional positional arguments for nn.Module.to().

• **kwargs – Additional keyword arguments for nn.Module.to().

Returns:

Self for method chaining.

reset_stats(initial_threshold: float | None = None)

Reset the running statistics.

Parameters:

initial_threshold – New initial threshold to use. If None, keeps the current threshold.

forward(x: Tensor, reset: bool = False, percentile: float = 50.0, *args: Any, **kwargs: Any) → Tensor

Apply dynamic thresholding to convert soft bit values to hard decisions.

Parameters:

• x – Input tensor of soft bit values.

• reset – If True, reset running statistics.

• percentile – Percentile to use if adaptation_method is ‘percentile’.

• *args – Additional positional arguments (unused).

• **kwargs – Additional keyword arguments (unused).

Returns:

Tensor of hard bit decisions (0.0 or 1.0).