bitorch.quantizations.dorefa.InputDoReFaFunction

class bitorch.quantizations.dorefa.InputDoReFaFunction(*args, **kwargs)

Methods

backward	just passes the unchanged output gradient as input gradient (i.e.
forward	quantizes input tensor and forwards it.

Attributes

static backward(ctx: Any, output_gradient: Tensor) → Tuple[Tensor, None]

just passes the unchanged output gradient as input gradient (i.e. applies straight through estimator)

Parameters:

• ctx (Any) – autograd context

• output_gradient (torch.Tensor) – output gradient

Returns:

the unchanged output gradient

Return type:

torch.Tensor

static forward(ctx: BackwardCFunction, input_tensor: Tensor, bits: int) → Tensor

quantizes input tensor and forwards it.

Parameters:

• ctx (Any) – autograd context

• input_tensor (torch.Tensor) – input tensor

• bits (int) – number of bits to round the input tensor to

Returns:

the quantized input tensor

Return type:

torch.Tensor