Method backward

src/bert_padding.py:42–53

(ctx, grad_output: torch.Tensor)

    @staticmethod
    def backward(ctx, grad_output: torch.Tensor) -> Tuple[torch.Tensor, None]:
        (indices,) = ctx.saved_tensors
        assert grad_output.ndim >= 2
        other_shape = grad_output.shape[1:]
        grad_output = rearrange(grad_output, "b ... -> b (...)")
        grad_input = torch.zeros(
            [ctx.first_axis_dim, grad_output.shape[1]], device=grad_output.device, dtype=grad_output.dtype
        )
        # TD [2022-03-04] For some reason torch.scatter is a bit faster than indexing.
        # grad_input[indices] = grad_output
        grad_input.scatter_(0, repeat(indices, "z -> z d", d=grad_output.shape[1]), grad_output)
        return grad_input.reshape(ctx.first_axis_dim, *other_shape), None


index_first_axis = IndexFirstAxis.apply