torch.repeatTensor() doesn't work with non-contiguous Tensor

[Atcold]

Let's sat I have a column from an image which I'd like to replicate horizontally. Only the first two row are computed correctly.

> = l
(1,.,.) = 
  0.6902
  0.6863
  0.6863
  0.6510
  0.6392

(2,.,.) = 
  0.6627
  0.6510
  0.6314
  0.5922
  0.5882

(3,.,.) = 
  0.6157
  0.5922
  0.5882
  0.5765
  0.5725
[torch.DoubleTensor of dimension 3x5x1]

> = l:repeatTensor(1,1,2)
(1,.,.) = 
  0.6902  0.6902
  0.6863  0.6863
  0.6627  0.6627
  0.6588  0.6588
  0.6471  0.6471

(2,.,.) = 
  0.6314  0.6314
  0.6157  0.6157
  0.6000  0.6000
  0.5765  0.5765
  0.5686  0.5686

(3,.,.) = 
  0.5569  0.5569
  0.5451  0.5451
  0.5176  0.5176
  0.5176  0.5176
  0.4980  0.4980
[torch.DoubleTensor of dimension 3x5x2]

[soumith]

is the matrix here contiguous? Can you post a code snippet to reproduce the issue?

[Atcold]

I cropped the first column of an image, therefore no, data are not contiguous.

require 'image'
lena = image.lena()
lenaColumn = lena[{ {},{1,5},{1} }]
= lenaColumn
= lenaColumn:repeatTensor(1,1,2)

which outputs

(1,.,.) = 
  0.7569
  0.7686
  0.7804
  0.7843
  0.7843

(2,.,.) = 
  0.4353
  0.4471
  0.4510
  0.4549
  0.4549

(3,.,.) = 
  0.2902
  0.3020
  0.3059
  0.3137
  0.3216
[torch.DoubleTensor of dimension 3x5x1]


(1,.,.) = 
  0.7569  0.7569
  0.7765  0.7765
  0.7647  0.7647
  0.7647  0.7647
  0.7843  0.7843

(2,.,.) = 
  0.7882  0.7882
  0.7804  0.7804
  0.7922  0.7922
  0.7725  0.7725
  0.7765  0.7765

(3,.,.) = 
  0.7686  0.7686
  0.7529  0.7529
  0.7529  0.7529
  0.7569  0.7569
  0.7451  0.7451
[torch.DoubleTensor of dimension 3x5x2]

[nicholas-leonard]

Yeah its definitely a problem of contiguity. When the input is contiguous, repeatTensor works.
Maybe we should just as an assert(tensor:isContiguous()) for now. Also, while we are in this function, we should probably allow passing a result tensor as is the case for most methods (allows for reusing memory).

[soumith]

closed via c19c892

[Atcold]

Thank you!