@alexpeys You could add a custom torch.autograd.Function (http://pytorch.org/docs/master/notes/extending.html#extending-torch-autograd) to compute the forward of matrix determinant, as well as the backward to obtain the gradients wrt each element of the matrix.

The codes would look like:

class Det(torch.autograd.Function):
    """
    Matrix determinant. Input should be a square matrix
    """

    @staticmethod
    def forward(ctx, x):
        output = torch.potrf(x).diag().prod()**2
        output = torch.Tensor([output]).cuda() # remove .cuda() if you only use cpu
        ctx.save_for_backward(x, output)
        return output

    @staticmethod
    def backward(ctx, grad_output):
        x, output = ctx.saved_variables
        grad_input = None

        if ctx.needs_input_grad[0]:
            grad_input = grad_output * output * x.inverse().t()

        return grad_input

At test time, just use Det.apply() in the computational graph.

To understand how the gradients are computed, please have a look at https://en.wikipedia.org/wiki/Determinant#Derivative

Can I plz has determinant function? #3423

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions