Pytorch in-place concatenation and conversion to tensor from numpy

Question

Suppose I have a list of tensors of the same size which could be concatenated along a dimension, say 0. Do any of the commands torch.cat or torch.stack or any numpy commands do the concatenation in-place? Also, if I want to convert a numpy ndarray to tensor. If I do the following, are two copies existing in the memory at any given time? I am dealing with a massive dataset so big only one copy of it can exist in the memory at any given time.

# initially data is a huge ndarray
data = torch.Tensor(data)

Answer 1

From your comment, assuming that:

1. you want to do something in the spirit of: B = A + a + b + ... + z where + represents concatenation along a compatible axis, B and A are huge and a , b , etc., are comparatively small, and
2. you can predict a reasonnable upper bound for the size of B

I would allocate beforehand a huge array for B using np.empty , and I would fill-in this array directly with your data as needed.