bitorch.layers.qlinear.QLinearBase

class bitorch.layers.qlinear.QLinearBase(*args: int, input_quantization: Optional[Union[Quantization, str]] = None, gradient_cancellation_threshold: Optional[float] = None, weight_quantization: Optional[Union[Quantization, str]] = None, **kwargs: bool)

Methods

__init__	Applies the given quantization functions on weights and inputs before applying the linear operation.
forward	Forwards x through the binary linear layer.
get_args_as_kwargs	Gather all arguments that were used to create a QLinear layer with argument names.

Attributes

gradient_cancellation_threshold
input_quantization

__init__(*args: int, input_quantization: Optional[Union[Quantization, str]] = None, gradient_cancellation_threshold: Optional[float] = None, weight_quantization: Optional[Union[Quantization, str]] = None, **kwargs: bool) → None

Applies the given quantization functions on weights and inputs before applying the linear operation.

Parameters:

• *args – positional arguments for linear layer

• input_quantization (Union[str, Quantization], optional) – quantization module used for input quantization. Defaults to None.

• gradient_cancellation_threshold (Union[float, None], optional) – threshold for input gradient cancellation. disabled if threshold is None. Defaults to None.

• weight_quantization (Union[str, Quantization], optional) – quantization module or name of quantization function. Defaults to None.

• **kwargs – keyword arguments for linear layer

forward(x: Tensor) → Tensor

Forwards x through the binary linear layer.

Parameters:

x (torch.Tensor) – tensor to forward

Returns:

forwarded tensor

Return type:

torch.Tensors

static get_args_as_kwargs(recipe: LayerRecipe) → Dict[str, Any]

Gather all arguments that were used to create a QLinear layer with argument names. Can be used to recreate a layer with identical arguments.

Returns:

A dictionary with all arguments (key is the argument name as a string even for positional arguments)