def backward()

in modules.py [0:0]

• 17 lines of code
• 3 McCabe index (conditional complexity)

    def backward(ctx, grad_output):

        # unpack saved context:
        input, weight, bias = ctx.saved_tensors

        # compute gradient with respect to input (not private):
        grad_input = grad_output.mm(weight)

        # compute norms of per-example gradients with respect to parameters:
        gradient_norm = grad_output.pow(2.).sum(1, keepdim=True).mul(
            input.pow(2.).sum(1, keepdim=True)
        )
        gradient_norm = torch.sqrt(gradient_norm)

        # aggregate the clipped per-example gradients:
        multiplier = _get_multipliers(gradient_norm, ctx.clip)
        grad_weight = grad_output.div(multiplier + 1e-9).t().matmul(input)

        # add noise to gradient:
        grad_weight += torch.randn_like(grad_weight) * ctx.clip * ctx.std

        # perform same procedure for bias gradients:
        if bias is not None:
            multiplier = _get_multipliers(grad_output.norm(2, 1), ctx.clip)
            grad_bias = grad_output.mul(multiplier.unsqueeze(1)).sum(0)
            grad_bias += torch.randn_like(grad_bias) * ctx.clip * ctx.std
        else:
            grad_bias = None

        # return private gradients:
        return grad_input, grad_weight, grad_bias, None, None