Will muti-level bracket be supported in rearragement in the future version?

[wzm2256]

Hi again,
while I am doing tensor multiplication, I always need to use bracket rearrangement. (One possible place these brackets appear is kronecker product as showed in my other discussion #271)

import torch
import einops
A=torch.randn(1, 32,32)

B = einops.rearrange(A, 'a b c -> (a (b c))')

#EinopsError:  Error while processing rearrange-reduction pattern "a b c -> (a (b c))".
#Input tensor shape: torch.Size([32, 32]). Additional info: {}.
#Axis composition is one-level (brackets inside brackets not allowed)

# I have to write it this way
B = einops.rearrange(einops.rearrange(A, 'a b c -> a (b c)'), 'a d -> (a d)')

In this example, I have to use two rearrangement twice, and I need to use one more index 'd', which is not intuitive. In practice, the situation can be much more complicated.

[boeddeker]

Hi,

you could use a single bracket.
There is no difference between (a b c), (a (b c)) and ((a b) c).
Only the letter order is important, not the execution order.

import torch
import einops
A=torch.randn(1, 32,32)
B = einops.rearrange(einops.rearrange(A, 'a b c -> a (b c)'), 'a d -> (a d)')
C = einops.rearrange(A, 'a b c -> (a b c)')
torch.testing.assert_close(B, C)

[wzm2256]

Oh, I see. The inner bracket can be deleted without infuencing the overall compuatation. That's cool. Thanks!