Unable to get the same values when converting from torch::Tensor to cv::Mat

C++
1

I am trying to convert a Torch-tensor to OpenCV-Mat int32

torch::Tensor A = torch::randint(0,30,{4,7},torch::TensorOptions().dtype(torch::kInt32));
cv::Mat cv_A(4,7, CV_32SC1);
std::memcpy((void*)cv_A.data, A.data_ptr(),sizeof(torch::kInt32)*A.numel());

When I print out the results, it shows

# Torch Tensor
 16  29   8  18   3  17  13
  4  17   0  26  22  21  26
 28  29  28  11  25  19  28
 11  25  13  11  23  20  28
[ CPUIntType{4,7} ]

and 
#OpenCV Mat
[16, 29, 8, 18, 3, 17, 13;
 0, 1, 4, 0, 0, 0, 0;
 0, 0, 0, 0, 0, 0, 0;
 0, 0, 0, 0, 0, 0, 0]

The results will not be same in-case of int-16.

torch::Tensor A = torch::randint(0,30,{4,7},torch::TensorOptions().dtype(torch::kInt16)); 
cv::Mat cv_A(4,7, CV_16SC1);
std::memcpy((void*)cv_A.data, A.data_ptr(),sizeof(torch::kInt16)*A.numel());

When printed the results for A and cv_A for the above code, I get

#torch Tensor
 6   6  10   3  26  29   3
  2  24  21  10  27   3   8
 21   8   8  27  19  18  22
 16   2   5   3  28  11  16
[ CPUShortType{4,7} ]

#opencv Mat
[6, 6, 10, 3, 26, 29, 3;
 2, 24, 21, 10, 27, 3, 8;
 0, 0, 0, 0, 0, 0, -26544;
 28670, 21872, 0, 0, 0, 0, 0]

However, I get the same results when uint8 is used.

 torch::Tensor A = torch::randint(0,30,{4,7},torch::TensorOptions().dtype(torch::kUInt8)); 
 cv::Mat cv_A(4,7, CV_8UC1);
 std::memcpy((void*)cv_A.data, A.data_ptr(),sizeof(torch::kUInt8)*A.numel());

#torch Tensor
20  21  22   5  10  21   1
  7   5  22  25  23  16   2
  8   3  15  20  17  16  27
 19  29  25   8   8  26  16
[ CPUByteType{4,7} ]

# OpenCV Mat
[ 20,  21,  22,   5,  10,  21,   1;
   7,   5,  22,  25,  23,  16,   2;
   8,   3,  15,  20,  17,  16,  27;
  19,  29,  25,   8,   8,  26,  16]

I am using the latest version of Libtorch ie 1.5.1
Can I know what is the issue here ?
Is there something wrong in the datatypes being used ?
P.S I have referred most of the issues posted in the forum and couldn’t come up with a solution for this.

1 Like
2

Found a work-around for this from here
However, the tensor won’t be copied.

cv::Mat cv_A(4, 7, CV_32SC1, A.data_ptr<int32_t>());

in the case of int-32 datatype.

3

This isn’t what you think it is (it is the size of a value of the ScalarType type), you probably want elementSize(torch::Int32).

Best regards

Thomas

2 Likes