I am starting to experiment with Pytorch C++ frontend in my C++ application.
My data is held in data arrays in my application and can easily be stored as DLTensors.
Is there a no copy way of initializing Pytorch Tensors using DLTensors or data arrays.
The only method I can find is torch::from_blob and for this, I need to clone the data to use it on GPU’s
Ideally, I’m looking for a no copy way to initialize tensors, like move semantics.