Initialising a complex tensor

Hi guys.

I’m creating a rotation matrix as a pytorch c++ tensor:

Where theta needs to be another tensor, which has requires_grad=true.

What would be the best practice way of initialising this tensor given the theta tensor? Do I just set each element manually, or is there some more clever way?

Cheers.

I’m clearly not using the right method. As a test I have:

    torch::Tensor theta = torch::tensor({0.18}, torch::dtype(torch::kFloat32).requires_grad(true));

    torch::Tensor cos_theta = torch::cos(theta);

    torch::Tensor t = torch::eye(4, torch::dtype(torch::kFloat32).requires_grad(false));

    t.index({0,0}) = cos_theta;

This produces an error:

output with shape [] doesn't match the broadcast shape [1]

And doesn’t seem like the right way of doing it either. Eventually I want to set each element of the tensor matrix individually with other tensors, such as:

t.index({0,0}) = cos_theta + (ux * (ux * one_minus_cos_theta));

t.index({0,1}) = (ux * uy * one_minus_cos_theta) - (uz * sin_theta);

t.index({0,2}) = (ux * uz * one_minus_cos_theta) + (uy * sin_theta);

//

t.index({1,0}) = (uy * ux * one_minus_cos_theta) + (uz * sin_theta);

t.index({1,1}) = cos_theta + (uy * uy * one_minus_cos_theta);

t.index({1,2}) = (uy * uz * one_minus_cos_theta) - (ux * sin_theta);

//

t.index({2,0}) = (uz * ux * one_minus_cos_theta) - (uy * sin_theta);

t.index({2,1}) = (uz * uy * one_minus_cos_theta) + (ux * sin_theta);

t.index({2,2}) = cos_theta + (uz * uz * one_minus_cos_theta);

I think I’m probably missing something conceptually with how to construct a tensor matrix out of individual tensor calculations - can someone assist?

Ah, it looks like the problem is with using .index. if I use t[0] it at least doesn’t crash.

I’m still interested in feedback on whether this is how you would approach the problem though.