如何循环 C++ 中 Pytorch 张量中的每个值？

Question

I am trying to do something like [0 if i < 1 else 1] for PyTorch tensor in C++. I tried to use tensor.accessor(), but it seem like it require you to know the dimension beforehand. While I want to pass it dynamically.

Is there anyway I could do this in C++ for Pytorch?

Answer 1

See if this helps:

Using TensorIterator:

https://labs.quansight.org/blog/2020/04/pytorch-tensoriterator-internals/

Or leveraging t.is_contiguous() / t.contiguous() to simplify transversal:

https://discuss.pytorch.org/t/iterating-over-tensor-in-c/60333/2

Answer 2

For loop in c++ torch is like this. Also, you can try torch.where. Be advised that for loop might be slow compared to built-in operations.

        auto answer = torch::zeros_like(x);
        auto batchCount = x.sizes()[0];
        auto pointCount = x.sizes()[1];
        
        // auto nextPoint = torch::zeros({batchCount, _pointDim});
        for (int i = 0; i < batchCount; ++i)
        {
            for (int j = 1; j < _pointCount; ++j)
            {
                answer[i][j] = answer[i][j - 1] + x[i][j - 1];
            } 
        }