[ 仅 API 调用方式不一致 ]torch.autograd.function.FunctionCtx.mark_non_differentiable

torch.autograd.function.FunctionCtx.mark_non_differentiable(*args)

paddle.autograd.PyLayerContext.mark_non_differentiable(*args)

# PyTorch 写法
class cus_func(torch.autograd.Function):
    @staticmethod
    def forward(ctx, x):
        a = x + x
        b = x + x + x
        ctx.mark_non_differentiable(a)
        return a, b

    @staticmethod
    def backward(ctx, grad_a, grad_b):
        grad_x = 3 * grad_b
        return grad_x

# Paddle 写法
class cus_func(paddle.autograd.PyLayer):
    @staticmethod
    def forward(ctx, x):
        a = x + x
        b = x + x + x
        ctx.mark_non_differentiable(a)
        return a, b

    @staticmethod
    def backward(ctx, grad_a, grad_b):
        grad_x = 3 * grad_b
        return grad_x