3 ms·
one step in the right directions are labeled dimensions, where you can write input = width * hight * color * batch and never have to worry about tensor permutat
by singularity2001 4y ago
one step in the right directions are labeled dimensions, where you can write input = width * hight * color * batch and never have to worry about tensor permutations again.
experimental support in pytorch
- amelius 4y agoIs that part of the type system?
- singularity2001 4y agomaybe I misremembered; https://pytorch.org/docs/stable/named_tensor.html https://pytorch.org/docs/stable/named_tensor.html
- albertzeyer 4y agoIt's called named dimensions, named tensors, dimension tags: https://pytorch.org/docs/stable/named_tensor.html https://pytorch.org/docs/stable/named_tensor.html It's already there in PyTorch since a while (2019). However, as far as I know, it has not found too much adoption yet. In our framework RETURNN, we also make heavy use of dimension tags: https://returnn.readthedocs.io/en/latest/getting_started/data.html https://returnn.readthedocs.io/en/latest/getting_started/dat...
- CoastalCoder 4y agoI forget the details, but I think Nervana also had a library(?) that used this concept before they were bought by Intel.