GridCellEncoder problems

[Thanh-Binh]

Hi Martin,
this week I have time to look at your codes, and found somethings in "GridCell2d.hpp" that seem to be not-correct (sorry, if I am wrong!):

1. The transformation matrix should not be inverted like

std::array<float, 4> mat = inv2x2({cosf(theta), -sinf(theta)
					,sinf(theta), cosf(theta)});

The correct formular should be:

std::array<float, 4> mat = {cosf(theta), -sinf(theta)
					,sinf(theta), cosf(theta)};

2. you use a utils- function
   inline Tensor gridCell2d()
   for encoding the different 2d-position values of the same input source.
   But every call this utils function creats new GridCell modules with different scale/rotation etc, so that the encoding results may not be consistant (i.e. encoding results of the same position are different from call to call

What do you think? Thanks

[marty1885]

Long time! How's going on your side? New stuff?

1. I think you are right. I vaguely remember there was a reason to use a inverse matrix (IIRC some semantic reason, not mathematical). But I can't recall now. Let me do some experiments and see if it there's a difference. But mathematically it should change nothing since inv(rotate(theta)) is just rotate(-theta). Thank you

2. No, every call to gridCell2d() with the same parameters will result in the same tensor. The scale/rotation of the modules are controlled by the seed parameter and has a default value of 42. As long as you didn't change seed, the encoding result shouldn't change.

auto t1 = encoder::gridCell2d({2, 2});
auto t2 = encoder::gridCell2d({2, 2});
assert((t1 == t2).all() == true);

[Thanh-Binh]

for 1. theoretically OK, but practically?
for 2.: OK