How to detect NaN during running?

[Sungyong Seo]

Hi guys

I am working on gaussian mixture model for unsupervised classifying.
BTW, during the calculation, certain points show 'nan' value.
So, I'd like to make them zero to prohibit error occurrence.
I tried to use 'isnan' but it doesn't work.

Do you know any way to filter 'nan' value?

Thanks,

[Alfredo Canziani]

Let ws be the Tensor of your weights,

-- Detecting and removing NaNs
if ws:ne(ws):sum() > 0 then
   print(sys.COLORS.red .. m.text .. ' weights has NaN/s')
   NaNOk = false
end
ws[ws:ne(ws)] = 0

[Sid]

I am trying to remove rows with NaN's altogether rather than set NaN to 0, here's what I've put together. Please let me know if there's a simpler way.

th> a = torch.linspace(1, 30, 30):reshape(10,3)      -- create a sample 2D tensor for the illustration

th> a
  1   2   3
  4   5   6
  7   8   9
 10  11  12
 13  14  15
 16  17  18
 19  20  21
 22  23  24
 25  26  27
 28  29  30
[torch.FloatTensor of dimension 10x3]

th> a[{{3,4},{2}}] = 0/0                            -- set a couple of values to nan for the illustration

th> a
  1.0000   2.0000   3.0000
  4.0000   5.0000   6.0000
  7.0000      nan   9.0000
 10.0000      nan  12.0000
 13.0000  14.0000  15.0000
 16.0000  17.0000  18.0000
 19.0000  20.0000  21.0000
 22.0000  23.0000  24.0000
 25.0000  26.0000  27.0000
 28.0000  29.0000  30.0000
[torch.FloatTensor of dimension 10x3]

th> nans = a:ne(a):max(2)                           -- find indices of nan and get the max along rows to get row indices

th> nans

 0

 0

 1

 1

 0

 0

 0

 0

 0

 0

[torch.ByteTensor of dimension 10x1]

th> indices = torch.linspace(1, a:size(1), a:size(1)):long()    -- a hacky step (currently unavoidable in torch)

th> indices

  1

  2

  3

  4

  5

  6

  7

  8

  9

 10

[torch.LongTensor of dimension 10]

th> a_clean = a:index(1, indices[nans:eq(0)])        -- pick the rows that don't have nan

th> a_clean

  1   2   3

  4   5   6

 13  14  15

 16  17  18

 19  20  21

 22  23  24

 25  26  27

 28  29  30

[torch.FloatTensor of dimension 8x3]

[Alfredo Canziani]

I’m not sure I can help…
But, if got it correctly, you want to “kill” rows in a Tensor, right?
Not sure this is “allowed”… nor if there is a smarter way then the one you showed.

​

[soumith]

Sid, 

That looks about right.

--
You received this message because you are subscribed to the Google Groups "torch7" group.
To unsubscribe from this group and stop receiving emails from it, send an email to torch7+un...@googlegroups.com.
To post to this group, send email to tor...@googlegroups.com.
Visit this group at http://groups.google.com/group/torch7.
For more options, visit https://groups.google.com/d/optout.

[Sid]

Alfredo, 

You're correct, that's what I'd like to do.