Implement tensor axis labeling for DIANNA completely about dianna HOT 8 CLOSED

The Pytorch named tensor looks interesting. I see a few pros/cons w.r.t xarray:

• pro: no need to specify a name for all columns
• reordering columns seems easier (moving channels is even an explicit example)
• con: we would need to depend on a very big library for just a small feature
• con(?): It is a prototype feature, less mature than xarray. The API may change

from dianna.

I like this a lot and it doesn't look like a huge amount of work. I think the reduction in complexity for the user is worth it.

from dianna.

I think this could be its own package/repo as it has little to do with DIANNA other than that we want to use it :-)

We could include some checks that give warnings if the number of channels for an image is not 1 or 3 etc.

from dianna.

Agreed, could be a spin-off. Just wanted to put the idea out there.

from dianna.

Maybe good to look into xarray.

from dianna.

I played around with xarray and implemented something in the xarray branch as a test for #55.

from dianna.

I came across this as well: https://pytorch.org/docs/stable/named_tensor.html

From what I read, this is not universally supported even within PyTorch itself (let alone outside, of course). Still, may be worth looking into how much exactly and how it compares to xarray.

from dianna.

Would be nice if conversion between xarray and NamedTensor were possible and easy.

from dianna.