torch.log() does not accept number as arguments

[yongfei25]

Hi, when trying to do util.logSoftMax(Z), got

/home/ubuntu/lib/torch/install/bin/luajit: invalid arguments: number 
expected arguments: [*CudaTensor*] CudaTensor
stack traceback:
    [C]: at 0x7f24a87612e0
    [C]: in function 'fn'
    ...all/share/lua/5.1/autograd/runtime/direct/DirectTape.lua:51: in function 'log'
    ...ubuntu/lib/torch/install/share/lua/5.1/autograd/util.lua:77: in function 'logSumExp'
    ...ubuntu/lib/torch/install/share/lua/5.1/autograd/util.lua:81: in function

(it worked with torch.setdefaulttensortype('torch.FloatTensor'), but failed with torch.setdefaulttensortype('torch.CudaTensor'))

[alexbw]

Can you provide a small snippet of runnable code to reproduce?

[yongfei25]

Hey @alexbw,

try running this code with CudaTensor and FloatTensor

local autograd = require 'autograd'

require 'cutorch'

torch.setdefaulttensortype('torch.CudaTensor')

local f = function(p)
  local x = torch.sum(torch.zeros(1,30))
  print(x)
  local y = torch.log(x) -- error
end

local df = autograd(f, { inline = false })
df(params)

(updated)

[yongfei25]

I updated the code above.

• It seems that torch.sum() returned a number when no dimension were specified, which is a correct behavior I think.
• The error occurred when the number is passed into torch.log() which only accepts Tensor as it's argument.
• It worked in FloatTensor, but not CudaTensor