bitorch.layers.qactivation.GradientCancellation

class bitorch.layers.qactivation.GradientCancellation(*args, **kwargs)

Methods

backward	Apply straight through estimator.
forward	Binarize input tensor using the _sign function.

Attributes

static backward(ctx: BackwardCFunction, output_grad: Tensor) → Tuple[Tensor, None]

Apply straight through estimator.

This passes the output gradient towards the input if the inputs are in the range [-1, 1].

Parameters:

• ctx (gradient context) – context

• output_grad (toch.Tensor) – the tensor containing the output gradient

Returns:

the input gradient (= the masked output gradient)

Return type:

torch.Tensor

static forward(ctx: BackwardCFunction, input_tensor: Tensor, threshold: float) → Tensor

Binarize input tensor using the _sign function.

Parameters:

input_tensor (tensor) – the input values to the Sign function

Returns:

binarized input tensor

Return type:

tensor