Create a tensor using torch::from_blob failed

C++
1

I use torch::from_blob in C++ to create a tensor, and failed with "The deleter and context arguments are mutually exclusive.“
it seems that the it is because the tensorMaker is with both deleter and context which i don’t know why.
my code is like below:
‘’’
x = torch::from_blob(data, {shape0, shape1}, torch::TensoorOptions().dtype(torch::kF32));
‘’’
p.s. it seems the error occurs when i update torch from 1.10 to 1.13. And only when building with release will this happen
i’d be grateful if anyone can give some advice

2

Could you try to create the tensor first using torch::from_blob, clone it afterwards, and finally transform to the desired dtype? torch::from_blob will share the same memory coming from data and in previous versions maybe providing additional options would create a copy of the data for you while it’s now disallowed.

4

hanks for reply!
in the code above, data is from a vector<float>, so i use tensorOptions.dtype(torch::kF32) not to change the dtype of the data but to specify the dtype of the data. i suppose this will not lead to copy of the data.
additionally, i tried your advice, but got the same failure, i’m not sure if i’ve done the correct thing as you suggest.

x = torch.from_blob(data, {shape0, shape1});
xc = x.clone();
xc.toType(torch::kI64);
5

supprise to find only calling from_blob(void*, at::IntArrayRef, const at::TensorOption &) alse leads to the same failure. code is like:

std::vector<float> data(shape0 * shape1 * 10, 1); // only to ensure the data size is enough
from_blob(data.data(), {shape0, shape1});
6

Ah OK, in this case I’ve understood the error message wrong as the at::TensorOptions argument it used to specify how to interpret the data. The error is raised from here](pytorch/Functions.cpp at fbc803df0c420db84429c51599f4fa4354b4493f · pytorch/pytorch · GitHub) but I’m unsure how both arguments are set.
Are you seeing the same issue using e.g.:

float data[] = {1, 2, 3, 4, 5, 6};
auto f = from_blob(data, {1, 2, 3});
8

I’ve found that the error is because that the directory of build is not cleared :joy: , something( perhaps cmake cache etc.) generated when i used the old version of libtorch is still there. thank you anyway.