( lazyvec1 - lazyvec1.t() ) @ lazyvec2 fails with: 'NoneType' object has no attribute 'replace'

[tvercaut]

Hi,

Thanks for the nice work. I am not sure if I am using the library correctly as it is the first time I try it but the following behaviour looks counter-intuitive to me. Can someone tell me if it's a bug or a misunderstanding from my part?

For two vectors u and v of size (Nx1) I would like to compute (u_l - u_l.T) @ v in a lazy fashion. Doing this directly with u_l being the lazy vector corresponding to the concrete vector u, and v being a concrete vector fails with the following error:

'NoneType' object has no attribute 'replace'

The workaround I found was to create a second lazy vector from the transposed concrete vector u.t() however this also was counter-intuitive since LazyTensor(u.t(),axis=1) is reported to be of size 1x1xN rather than 1xN. The second workaround I found was to create the lazy vector from a 3D tensor of size 1xNx1.

Let me know what I might be missing.

Here is an example (also on colab):

import torch
print(torch.__version__)

torchdevice = torch.device('cpu')
if torch.cuda.is_available():
  torchdevice = torch.device('cuda')
  print('Default GPU is ' + torch.cuda.get_device_name(torch.device('cuda')))
print('Running on ' + str(torchdevice))

!pip install pykeops[colab]
import pykeops
print('pykeops version:',pykeops.__version__)

from pykeops.torch import LazyTensor

# Generate 2 vector u and v
u = torch.normal(0, 1, size=(5,1), device=torchdevice)
v = torch.normal(0, 1, size=(5,), device=torchdevice)

# Compute (u - u.T) @ v in a dense fashion
P = u - u.t()
Pv = P @ v
print('Pv: ',Pv)

# Compute (u - u.T) @ v in a lazy fashion
u_l = LazyTensor(u,axis=0)
print('u_l: ', u_l)

# Using the transpose of the lazy tensor doesn't work here
#print('u_l.t(): ', u_l.t())
#P_l =  u_l - u_l.t()
#print('P_l:', P_l)

# We thus go for transposing the concrete vector and creating a new lazy tensor
# However, it seems to fails when provided as a 2D matrix of size 1 x N
#ut_l = LazyTensor(u.t(),axis=1)
print('LazyTensor(u.t(),axis=1): ',LazyTensor(u.t(),axis=1))
ut_l = LazyTensor(u.t()[...,None])
print('ut_l: ', ut_l)

P_l =  u_l - ut_l
print('P_l:', P_l)

Pv_froml = P_l @ v
print('Pv_froml: ',Pv_froml)

# Putting this at the end just to highlight that this fails
( u_l - u_l.t() ) @ v

[joanglaunes]

Hello @tvercaut ,
I have made this an issue because it is clearly a bug, I will answer you in the issue section