These are chat archives for torch/torch7

21st
May 2015
yuntang
@yuntang
May 21 2015 00:36 UTC
I install torch on a gpu server (K20). When I try to load cunn, torch will take over all 4 GPU cores on that server. If any core is used by other program, torch will fail to load cunn with error message below:
cuda runtime error : all CUDA-capable devices are busy or unavailable at /tmp/luarocks_cutorch-scm-1-9207/cutorch/lib/THC/THCGeneral.c:237
I am not sure if it is normal that torch will take all GPU resources by default. If that is so, how can we configure it to use one GPU card only
Soumith Chintala
@soumith
May 21 2015 01:25 UTC
@yuntang to see only GPU0, start torch with: CUDA_VISIBLE_DEVICES=0 th [yourscript]
Also, please open github issues if you want to ask question about error messages, because they will be helpful for future users (who can do a google search).
samuel
@samuel1208
May 21 2015 03:00 UTC
is anyone interest in "Delving Deepinto Rectifiers: Surpassing Human-Level Performanceon ImageNet Classification"
yuntang
@yuntang
May 21 2015 05:51 UTC
@soumith It works. Thanks.
Rémi Francis
@francisr
May 21 2015 09:35 UTC
@nicholas-leonard Thanks, I looked past these layers because I confused then with CMulTable and AddTable.
@soumith Is there a way to specify the GPUs used from inside a torch script (before loading cutorch)?
John Armstrong
@jarmstrong2
May 21 2015 15:06 UTC
anyone have any insight into why torch.DiskFile might not save a table to a file? Could it be a memory issue? also this table has chars as its indices
John Armstrong
@jarmstrong2
May 21 2015 15:17 UTC
oh nevermind just used torch.save(filename, tablevar) instead
John Armstrong
@jarmstrong2
May 21 2015 15:46 UTC
any recommendations for a good ide when working with lua/torch?
Soumith Chintala
@soumith
May 21 2015 15:54 UTC
@francisr not from inside torch. it has to be done before the process starts at the nvidia driver level. that's why the environment variables.
Ajay Talati
@AjayTalati
May 21 2015 19:08 UTC
Hey folks, @karpathy , @soumith or anyone else who's working on simultaneous language-image modelling, do you guys know of any sentence/caption and image datasets in other languages, perhaps multilingual? For example, anything like Flicker 8, 30 or even MS COCO in another language? Thank you :)
Bas Veeling
@basveeling
May 21 2015 19:47 UTC
Newbie question about tensor indexing; say I have a 1d tensor with positive values, and missing data is represented as 0.0. How can I quickly find the indexes where the value != 0.0, and how to use these indexes to select the these non-zero values from the tensor? Right now I have:
indexes = torch.LongTensor(torch.find(torch.ne(frequencies,0.0),1))
freqs = frequencies:index(1,indexes)
(using torchx for the torch.find function)
Bas Veeling
@basveeling
May 21 2015 20:05 UTC
(which seems somewhat unwieldy)
Alfredo Canziani
@Atcold
May 21 2015 20:10 UTC
@basveeling, I'm not sure if this is what you've asked for, but what about
freqs = frequences[frequences:gt(0)]
Alfredo Canziani
@Atcold
May 21 2015 20:16 UTC
For example
a = torch.range(1, 10)
a:apply(function(x) if x % 2 == 0 then return 0 end end)
print(a[a:gt(0)])
---
 1
 3
 5
 7
 9
[torch.DoubleTensor of dimension 5]
Alfredo Canziani
@Atcold
May 21 2015 20:32 UTC
@karpathy: http://karpathy.github.io/2015/05/21/rnn-effectiveness/ WOW! I haven't read it yet but it looks AMAZING!
Andrej
@karpathy
May 21 2015 20:48 UTC
@Atcold thanks! It takes a while to get through :)
Alfredo Canziani
@Atcold
May 21 2015 20:51 UTC
@karpathy :+1: