Question about your convexLS solver

[pramanc]

Hi, It's me again!

I met a problem when calculate the gradient ( solveP function) .

when calculate z by solve the linear system A * z = b

some times A is zero matrix-> cause the calculated z is nan.

I am reading the code of convexLS solver, but can't find any reference.

Would you please tell me some reference of the convexLS solver?

Thanks!

[binhhuyle]

Hi @pramanc,

A good reference that I would recommend is the book "Numerical Methods for Least Squares Problems" by Ake Bjorck.
You can check two chapters:

• Constrained Least Squares Problems --> Linear Equality Constraints
• Constrained Least Squares Problems --> Linear Inequality Constraints

I wrote a little bit (not very completed) about my implementation here: https://uh-ir.tdl.org/handle/10657/1824
(in chapter 3.7 Skinning Weights Update)

[binhhuyle]

Re: A is zero matrix problem

A in solveP function should not be zero. It is positive definite due to the construction, it is actually A^T*A (A transposed * A).

[pramanc]

@binhhuyle Thank you for your reply!

zero matrix problem actually happens in

solveP function:
z=q2[np-2]*( //Re-project
(q2[np-2].transpose()*indexing_row_col(aTa, idx.head(np), idx.head(np))q2[np-2]).colPivHouseholderQr().solve( //A
q2[np-2].transpose()(indexing_vector(aTb, idx.head(np))-indexing_row(aTa, idx.head(np))*x) )); //b

the matrix (q2[np-2].transpose()*indexing_row_col(aTa, idx.head(np), idx.head(np))*q2[np-2]) sometimes will be zero matrix

=======================================================
I will read the reference to see why it happens

Thank you again! Really amazing work!

[robertjoosten]

Hello, hope both of you are doing well.
I think I'm running into the same problem and wanted to double check to see if either of you have found a solution or the cause of the problem.

After solving p it contains nan values and zeros. Which results in iMin being -1.
_Scalar eps=std::abs(x(idx[iMin]));

Just debugged the values after the solve and I am doing something terribly wrong.
aTa:
-nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind)
-nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind)
-nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind)
-nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind)
-nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind)
aTb:
-nan(ind)
-nan(ind)
-nan(ind)
-nan(ind)
-nan(ind)
aTa:
-nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind)
-nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind)
-nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind)
-nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind)
-nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind)
aTb:
-nan(ind)
-nan(ind)
-nan(ind)
-nan(ind)
-nan(ind)
x:
0.876707
0
0
0.123293
0
np:
2
affine:
1
p:
-nan(ind)
0
0
-nan(ind)
0

[pramanc]

@robertjoosten Hi, Can you supply the data(aTa, aTb, x, idx, np, affine, p) passed in solveP?

[robertjoosten]

Hello @pramanc

I've narrowed it down a bit further. I think it goes wrong because of nan values introduced in m during the computeTransformations step.

m looks fine when it comes in ( identity matrix )
When computing vuT nan values make their way into the vuT which propagate to m ( I am assuming, in over my head with this level of math )

Here are some debug values for values in the SparseMatrix::InnerIterator in compute_vuT
The height it value coming from the weights makes me thing that might be where the problem lies.

it row
5
it value
6.15434e+255
v
0.987688
-0.156434
0
u
0.987688 -0.156434 0 1
temp
0.975528 -0.154508 0 0.987688
-0.154508 0.0244717 -0 -0.156434
0 -0 0 0
0.987688 -0.156434 0 1
vuT
6.00373e+255 -9.50897e+254 0 6.07857e+255
-9.50897e+254 1.50607e+254 -0 -9.6275e+254
0 -0 0 0
6.07857e+255 -9.6275e+254 0 6.15434e+255
vutP
inf -inf -nan(ind) inf
-inf inf -nan(ind) -inf
-nan(ind) -nan(ind) -nan(ind) -nan(ind)
inf -inf -nan(ind) inf

I really doubt this is related to the original issue anymore, happy to take the discussion away from this thread
Edit: this was me populating the weights incorrectly causing the matrices to flip out... sorry for the inconvenience

[pramanc]

6.15434e+255

well, definitely your sparse matrix w is wrong to make it value(6.15434e+255)，you can check the fbxreader part to see if you get right weight matrix w