Problems with forward in new style function

autograd
1

I am trying to implement custom function for gaussian CRF layer and weird error of forward() takes exactly 3 arguments (2 given) during apply()occurs. As I use new style static method based forward and backward, I have no idea what is basically wrong in current implementation, so can somebody point out the mistake.

The function code:

class G_CRF_layer(torch.autograd.Function):

    @staticmethod
    def forward(ctx, unary, pairwise):
        positive_mul = 1e-3
        decomposition = cho_factor(pairwise.numpy() + positive_mul * np.eye(pairwise.size(0)))
        x = cho_solve(decomposition, unary.numpy())
        
        x = torch.from_numpy(x).float()
        ctx.data_for_backward = pairwise, x, decomposition
        return x

    @staticmethod
    def backward(ctx, grad_output):
        positive_mul = 1e-3
        pairwise, x, decomposition, = ctx.data_for_backward

        unary_grad = torch.from_numpy(cho_solve(decomposition, grad_output.data.numpy())).float()
        pairwise_grad = (- unary_grad * x.view(1,-1))

        return Variable(unary_grad), Variable(pairwise_grad)

And the apply function call:

A, B = np.array([[1,1],[1,2]]), np.array([3,5])
pairwise = Variable(torch.from_numpy(A).double(), requires_grad=True)
unary = Variable(torch.from_numpy(B).double().view(-1,1), requires_grad=True)

input = unary, pairwise
G_CRF_layer.apply(input)
2

Well, your forward function takes both unary and pairwise as arguments. But you call apply with a single tuple argument.
You should change to G_CRF_layer.apply(*input).

3

Thank you for your quick response, I’ve probably missed that detail.