Resolved Quaternion * vector3 return NaN ?

Hello everyone,
I'm creating a neural networks where inputs and outputs are Vector3, and weights are Quaternions.
Long story short, Here is my code for the maths of the nn :

Code (CSharp):

1.         public Vector3 Proceed(Vector3[] _in)
2.         {
3.             Vector3 result = Vector3.zero;
4.  
5.             for(int i = 0; i < _in.Length; i++)
6.             {
7.                 result += rotations[i] * _in[i] * weights[i];
8.                 //NaN somewhere here... why ?
9.             }
10.  
11.             return result;
12.         }

rotations is an array of quaternions, and weights is an array of floats.
However, after this I have vector3 output that have component that is NaN... anyone knows where that come from ? I am doing something wrong ?

 

Check i your quaternion is defined and initiated. Meaning is none 0, or Nan matrix. You should have some values, or at least be identity.

I think also, your multiplied vector should also be none 0. Consider using normalised vector.

 

Thanks for the reply.
The quaternion is created from Quaternion.Euler(random vector3 between -180 and 180).
It works when I normalize the result of the multiplication of the Quaternion and the vector, but by doing this I am losing information in the neural network, and I need the magnitude of the vectors.

 

What I mean is, you could normalise vector before multiplication.

You shouldn't be normalising after multiplication, as indeed, you will loose information.

Can you you write here a matrix and vectors of one entry case, which results NaN?
For example every value for index i=0.

Also, what you can try, remove temporarily multiplication by vectors, and simplyultiply quaternion by forward vector. See if you got valid data, or NaN. At least will now, if quaterrnion is valid.

If is valid, then try remove / replace one multiplication parameter, with constant.
And see if results are valid.

 

Done a bit of tweaking.
When I normalize input vector before multiplication, everything works, But I still lose the information on the input vector length.
I saw that the input vector's magnitude was NaN, but only after some iterations throughts the neural network

 

Looks like this may be your issue.
You must not allow NaNs into system.

 

Yes, that was it. Add if conditions before feeding the system, works nicely now.
Thanks a lot for the help !