How does torch::where work in libtorch?

C++
1

Hello everyone,
I’m porting a python module into c++ and I’m stuck here as how to efficiently port :

np.where(np_array > some_float)

numpy returns two tuples of indexes. I found out torchlib has a where method as well but it returns a at::autograd::variable_list which I guess is simply a std::vector<torch::Variable> .
I couldnt find any related documentation as how to go about it and whether this equivalent to the np.where!
Running a simple test, shows that the return value of the torch::where seems to be having two tensors as the result. and each seem to represent a set of indexes.
However, Its not clear to me, how to interpret these numbers, since there are duplicate numbers and I cant make anything out of it!
Here is the simple demo :

#include <iostream>
#include <torch/torch.h>
#include <torch/script.h>

int main()
{
    torch::Tensor preds = torch::randn({ 4,8 });
    torch::Tensor offsets = torch::randn({ 1, 4, 4, 8 });
    auto res = torch::where(preds > 0.9);
    std::cout << "res0.size(): " << res[0].sizes()[0] << std::endl;
    if (res[0].sizes()[0] == 0)
    {
       std::cout << "nothing is returned!" << std::endl;
    }
    
    std::cout << "preds:\n" << preds << std::endl;
    std::cout << "offsets:\n" << offsets << std::endl;

    std::cout << "res.size: " << res.size() << std::endl;
    std::cout << "res0:\n" << res[0] << std::endl;
    std::cout << "res1:\n" << res[1] << std::endl;
    std::cout << "Hello wordl!" << std::endl;
    std::system("pause");
}

this produces this output:

res0.size(): 9
preds:
-0.9659 -0.9782 -1.7985 -1.5461 -0.2913  0.2766 -1.6117 -0.4642
-1.8279  0.9883 -1.0430 -0.0729  1.4847 -0.8073 -1.8110  0.1578
 0.6299  0.6409  1.0483  0.9996 -0.5102 -0.5886  0.9410  0.3633
-0.4674  1.4806  1.0258 -0.5636  0.9242 -0.1481 -1.8079  2.4731
[ CPUFloatType{4,8} ]
offsets:
(1,1,.,.) =
 -0.1614  0.3842 -0.8396 -0.0390 -0.5792  0.1151 -1.4113  0.4774
  0.1490  1.7119 -0.3051 -0.1585 -1.2240 -0.7920 -0.7085 -0.1590
 -0.6625  1.6923 -0.0494 -0.5374 -0.7239 -0.6688 -1.7225  0.3280
  0.8866  2.4275 -0.4948  1.6180 -0.4133 -0.5967  0.3257 -0.7997

(1,2,.,.) =
 -0.5249  1.8908 -0.5924  1.3851  0.8556  0.4914  0.6360  0.8373
  0.2659 -1.3615  0.1923 -0.9496 -1.8631 -0.1948 -0.3652  0.0215
 -0.2501  2.2011  0.8395 -0.4594 -0.2585  0.1908  0.4444  0.2588
  0.5341 -1.6821  0.3597 -0.0394 -0.0743  0.5159 -0.9592  1.0634

(1,3,.,.) =
 -1.3950 -0.7348  1.1979  1.0431  0.4164 -0.1143  0.0959  0.1557
  2.0382  0.0820  0.2825  0.5211  2.0970  0.2202 -1.4359 -1.0877
  0.1951  0.4328 -0.2317 -0.7700 -0.4143 -1.8647 -0.6055  0.8088
  0.7052 -0.1012  0.3797 -1.1983 -1.0531  0.6381 -0.6158 -0.6769

(1,4,.,.) =
 -1.0985 -1.0719 -0.3173 -0.2694  1.4531 -0.7597  0.4511  0.5871
 -1.2546  1.6486 -1.1799  0.7037 -0.6020 -0.8686  1.3177 -1.1100
  0.4443 -0.7602 -1.3926 -0.0029 -0.9942  1.0145 -1.4573 -0.0027
  0.5276  0.5788  0.7841 -0.1466  0.3800 -0.0146 -0.1285 -2.1790
[ CPUFloatType{1,4,4,8} ]
res.size: 2
res0:
 1
 1
 2
 2
 2
 3
 3
 3
 3
[ CPULongType{9} ]
res1:
 1
 4
 2
 3
 6
 1
 2
 4
 7
[ CPULongType{9} ]
Hello wordl!

So any help is greatly appreciated

2

OK! I figured it out!
Each number indicates a value in that row that satisfies the condition.
for example here :

1
 1
 2
 2
 2
 3
 3
 3
 3

it means that in row 1, 2 elements satisfy the condition .in row 2, 3 elements satisfy the condition, and likely the row 3, 4 values satisfy the conditions. in the second part (i.e. res1) the actual indexes for those values are provided that is :

1
 4
 2
 3
 6
 1
 2
 4
 7

1 means, the first value in the row 1
4 means, the 4th value in the row 1
2 means the 2nd value in the row 2
3 means the third value in the row 2
and so on…

to visualize this, simply use torch::ones instead of randn and see the output like this :

res0.shape: 32
preds:
 1  1  1  1  1  1  1  1
 1  1  1  1  1  1  1  1
 1  1  1  1  1  1  1  1
 1  1  1  1  1  1  1  1
[ CPUFloatType{4,8} ]

res.size: 2
res0:
 0
 0
 0
 0
 0
 0
 0
 0
 1
 1
 1
 1
 1
 1
 1
 1
 2
 2
 2
 2
 2
 2
 2
 2
 3
 3
 3
 3
 3
 3
 3
 3
[ CPULongType{32} ]
res1:
 0
 1
 2
 3
 4
 5
 6
 7
 0
 1
 2
 3
 4
 5
 6
 7
 0
 1
 2
 3
 4
 5
 6
 7
 0
 1
 2
 3
 4
 5
 6
 7
[ CPULongType{32} ]

and it works just lik np.where