These are chat archives for torch/torch7

12th
Jun 2015
Rémi Francis
@francisr
Jun 12 2015 10:42
Has anyone used the SCRNN code from https://github.com/facebook/SCRNNs/blob/master/mfactory.lua ? It doesn't seem to follow the equations from the paper.
Namely the update of the slow layer is s = (1-a) * Bx + s instead of being s = (1-a) * Bx + a*s .
Praveer SINGH
@PraveerSINGH
Jun 12 2015 10:49
Is there an implementation in torch for Visualizing the Saliency maps like here : http://arxiv.org/pdf/1312.6034.pdf ?
Laurens van der Maaten
@lvdmaaten
Jun 12 2015 12:55
@PraveerSINGH Can't you just generate a 1xWxH tensor with numbers between 0 and 1 representing the saliency at that location, and export that tensor with image.save()?
Soumith Chintala
@soumith
Jun 12 2015 14:49
@francisr I invite you to write to Sumit Chopra at his email. He will definitely clarify your doubts..
Davide Chicco
@davidechicco
Jun 12 2015 16:08
Hi guys
loken
@lokenetwork
Jun 12 2015 16:09
Some people know DMLC form baidu?
Davide Chicco
@davidechicco
Jun 12 2015 16:09
I'm having problems with my siamsese neural network in Torch. My gradient update does not update anything :-( Could someone take a look to my code?
Alfredo Canziani
@Atcold
Jun 12 2015 16:11
Does anyone know how to check whether ImageMagick uses libjpeg-turbo8? That should be the default behaviour, but I'd rather be sure it does it.
Alfredo Canziani
@Atcold
Jun 12 2015 16:17
@PraveerSINGH, yes, I've implemented that and also the rest of my paper. I was planning on releasing it last week, but I got a tad busy. I'll do it ASAP. Ping me if I just lied.
Alfredo Canziani
@Atcold
Jun 12 2015 17:42
By lld /usr/lib/libGraphicsMagick.so I get libjpeg.so.8 and not the libjpeg-turbo.so.8... :disappointed_relieved:
Paweł Szczur
@orian
Jun 12 2015 17:42

i have a vector I want to add to all rows,

x = torch.Tensor(torch.LongStorage({K}))
y = torch.Tensor(torch.LongStorage({N,K}))
y:dimadd(1,x)  -- adds x to all elements by dimention 1

Is there a simple way to do that in torch?

Right now i have a code:

for i=1,N do
     y[i]:add(x)
end
Alfredo Canziani
@Atcold
Jun 12 2015 17:43
@orian, please use 3 back ticks for inserting code. Press UP for editing your message. Tnx.
For adding, just use x:add(dim).
Soumith Chintala
@soumith
Jun 12 2015 17:44
@Atcold LD_PRELOAD=[path-to-libjpeg-turbo] th yourscript.lua
Paweł Szczur
@orian
Jun 12 2015 17:46
@Atcold I've updated the question, simple add won't do that unless the tensors have same number of elements.
Davide Chicco
@davidechicco
Jun 12 2015 17:55
Please help me
Alfredo Canziani
@Atcold
Jun 12 2015 17:56
@orian, y:add(x:expandAs(y))
Paweł Szczur
@orian
Jun 12 2015 17:57

@Atcold: Oh, I didn't know about that, thanks! I've checked that doc, this is super useful: https://github.com/torch/torch7/blob/master/doc/tensor.md#result-expandresult-sizes

I had to first get a view of data with the same number of dimentions: y:add(x:view(1,K):expandAs(y))

Alfredo Canziani
@Atcold
Jun 12 2015 18:11
@orian :+1:
Alfredo Canziani
@Atcold
Jun 12 2015 19:27
@orian, actually, you can y:add(x:view(1,-1):expandAs(y)).
Sandy4321
@Sandy4321
Jun 12 2015 20:27
pls help to find way how to use visual debugger: I use zbs-torch and ZeroBrane Studio from https://github.com/torch/torch7/wiki/Cheatsheet, but images can not be created??
if somebody use http://www.eclipse.org/koneki/ldt/ LDT debugger for Torch to be able to do step by step debugging and create images coded in file and what is more important create new images to see what happen inside code?
Sandy4321
@Sandy4321
Jun 12 2015 20:33
pls help to debug line by line C++ or C code , which is used in Torch code? so I have Torch code uses some files written in C++, how can I go in C code, to see what happen in C code, to see value of variables to plot graphs from vector arrays?
Houshalter
@Houshalter
Jun 12 2015 22:15
how difficult would it be to modify char-rnn's lstm to add one neuron at a time, after it's trained for awhile?
Soumith Chintala
@soumith
Jun 12 2015 23:13
@Houshalter do you mean that you want to add more neurons to the LSTM after the existing ones have been trained for a while?
Houshalter
@Houshalter
Jun 12 2015 23:14
@soumith yes
@soumith but I don't know how the LSTM module works or if that's possible
i can't even figure out where the code is for the lstm module char-rnn is using. it requires it from "model.LSTM"
Soumith Chintala
@soumith
Jun 12 2015 23:19
@Houshalter it's in the model folder. It looks like you might want to get a little more familiar with torch/lua first before directly attacking the LSTM module written using nngraph. Doing a few simpler tutorials like https://github.com/soumith/cvpr2015/blob/master/Deep%20Learning%20with%20Torch.ipynb might be a good exercise.
Houshalter
@Houshalter
Jun 12 2015 23:20
oh i see it was requiring something in it's own folder. I didn't know you could do that.