einops.layers.torch.Rearrange does not accept a list[torch.Tensor] as an input

[zimka]

Describe the bug einops.layers.torch.Rearrange and einops.rearrange behave differently with list[torch.Tensor]. It is possible to use rearrange for concatenation, while Rearrange fails with error:

einops/_torch_specific.py:66, in TorchJitBackend.shape(x) 64 @staticmethod 65 def shape(x): ---> 66 return x.shape AttributeError: 'list' object has no attribute 'shape'

Reproduction steps

import torch
from einops import rearrange
from einops.layers.torch import Rearrange

layer = Rearrange('num b c h w -> b (num c) h w', num=2)
func = lambda x: rearrange(x, 'num b c h w -> b (num c) h w', num=2)

t1 = torch.randn((1, 3, 16, 24))
t2 = torch.rand((1, 3, 16, 24))

y = func([w1, w2]) # OK
y = layer([w1, w2]) #  AttributeError: 'list' object has no attribute 'shape'

Expected behavior Expect the same behavior in both cases (hopefully correct operation execution, not raise exception)

Your platform python==3.10.12 torch==2.0.1 einops==0.6.1

[arogozhnikov]

Yup, that's how it was supposed to work. Nobody asked for support of lists in layers ... yet :)