torch5-devel — News and discussions about Torch5 development

[Torch5-devel] Join my network on LinkedIn through Rainmaker

[Clément F. <cle...@gm...>]

LinkedIn
------------



I would like to connect with you on LinkedIn.

Powered by Rainmaker - Get your free account at http://rainmaker.cc/

Clément Farabet
Research Scientist at New York University
Greater New York City Area

Confirm that you know Clément Farabet:
https://www.linkedin.com/e/cn2tuy-gyl1mi5g-4f/isd/5896331344/rQSlaaP0/?hs=false&tok=3lqZPOInCl0B81

--
You are receiving Invitation to Connect emails. Click to unsubscribe:
http://www.linkedin.com/e/cn2tuy-gyl1mi5g-4f/hSAhDk1mDGRjO9Dtteq7sk3F8TBWBfYCjxAGvhJomTPWdc3C5Lt/goo/torch5-devel%40lists%2Esourceforge%2Enet/20061/I2050415880_1/?hs=false&tok=3TR4CUbLOl0B81

(c) 2012 LinkedIn Corporation. 2029 Stierlin Ct, Mountain View, CA 94043, USA.

[Torch5-devel] Memory leaks

[Clément F. <cle...@gm...>]

Hi all,

I also noticed some severe memory leaks (lack of garbage collection) on
tensor allocation.

Try to run:
> for i = 1,1000000 do a = torch.Tensor(500,500) end
(you can stop it before it consumes all your mem...)

at the end of this for loop, there's only one valid reference to a tensor,
but nothing ever gets garbage collected. Even a manual call to the garbage
collector doesn't clean that up.

Maybe I'm mistaken, and that kind of data cannot be handled by the garbage
collector ?
If it is so the work around is to always pre-allocate tensors, and only
resize/fill them with new data in loops, as you would do it in C... but that
sort of defeats the purpose of a scripting language right ?

Other (less important) leak, the function qt.QImage.fromTensor(tensor)
returns a freshly allocated QImage, that doesn't seem to be garbage
collectible. That basically makes QT displays have a limited lifetime :-(,
as any tensor needs to be converted like that first.
I've written an other function that fills a QImage passed as an argument, to
avoid this leak, but it would be better if the QImage was simply collected
at some point.

Clement.

[Torch5-devel] Compile torch5 from source, with none of 3rdparty libraries, how to?

[JIA P. <jp...@gm...>]

Hi, all:

Since all the 3rd party libraries can be grabbed from Ubuntu repository
directly,
I'm wondering how can I compile current torch5-SVN after removing the
subdirectory "3rdparty" ?
The reason why I'd love to do so is just because some 3rdparty libraries are
a bit old and
I seriously don't want to install two version of 3rdparty libraries in one
computer.
For example:

3rdparty folder                Ubuntu repository
lua-5.1.3                                  lua-5.1.4
cairo-1.6.4                               cairo-1.8.10
etc...

Can anybody help??? By the way, is torch5 still alive???


I'm currently using CMake and after removing the entire "3rdparty"
subfolder,
I obtained the following error messages:




 CMake Error at scripts/TorchPackage.cmake:29 (FIND_PACKAGE):
   Could not find module FindLua.cmake or a configuration file for package
   Lua.

   Adjust CMAKE_MODULE_PATH to find FindLua.cmake or set Lua_DIR to the
   directory containing a CMake configuration file for Lua.  The file will
   have one of the following names:

     LuaConfig.cmake
     lua-config.cmake

 Call Stack (most recent call first):
   CMakeLists.txt:52 (INCLUDE)



 CMake Error at libraries/luaT/CMakeLists.txt:3 (FIND_PACKAGE):
   Could not find module FindLua.cmake or a configuration file for package
   Lua.

   Adjust CMAKE_MODULE_PATH to find FindLua.cmake or set Lua_DIR to the
   directory containing a CMake configuration file for Lua.  The file will
   have one of the following names:

     LuaConfig.cmake
     lua-config.cmake




 CMake Error at libraries/TH/CMakeLists.txt:13 (CHECK_FUNCTION_EXISTS):
   Unknown CMake command "CHECK_FUNCTION_EXISTS".















-- 
Welcome to Vision Open
http://www.visionopen.com

[Torch5-devel] help

[sridhar i. <sri...@gm...>]

hai
I am devolping an activity recognition module using torch5(thanks a lot for
this wonderful env.!!) . I was taking the tut for nn's. I have declared the
datasets as per the tut. But everytime I issue the "trainer:train(dataset)"
command it throws this error
# StochasticGradient: training
...ogram Files/Torch5 0.8.0/share/lua/5.1/nn/Linear.lua:40: size mismatch
stack traceback:
[C]: in function 'addT2dotT1'
...ogram Files/Torch5 0.8.0/share/lua/5.1/nn/Linear.lua:40: in function
'forward'
...m Files/Torch5 0.8.0/share/lua/5.1/nn/Sequential.lua:26: in function
'forward'
...Torch5 0.8.0/share/lua/5.1/nn/StochasticGradient.lua:36: in function
'train'
C:/Program Files/Torch5 0.8.0/simptutnn:23: in main chunk
[C]: in function 'dofile'
[string "dofile('C:/Program Files/Torch5 0.8.0/simpt..."]:1: in main chunk
[C]: ?
|| finished after 0.359 seconds
-- 
kindly help!!

yours truly
Krishnan

-- 
Krishnan

[Torch5-devel] Memory leaks are consuming all computer memory

[Franck L <fra...@gm...>]

Hi all,

I am very new to Torch5 and Lua, but experienced in C and other NN libraries
(Fann, libann).
First congratulations for the work, I like the modular way of Torch.

I have been making some "tries" using some data previously used with Fann
library. Just few results and questions/comments:


   - I use a dataset with 38 input (normalized ramge -1 +1) and 1 output
   (binary 1 - 0). I have about 300 000 samples in the dataset
   - A first simple test with 1 nn.Linear(38, 3), and a Tanh function using
   an MSECriterion for the error consume about 3 Mb of rmemory per iteration !
   - I have used the trainer.train method and also built the "manual" way, I
   got the same results about memory leaks

I wounder where is the leak located. I admit I am very new to Lua, but I
guess this is located in the C code somewhere. Did you guys already noticed
this ?

Then few questions comparing Torch with other library, mainly the result of
the MSE error .

When I run the above described network (300 000 records, 38 input, 1 output)
using a learning rate of 0.01 and no "decay", the error just seems to be the
same after the second iteration. I can run about 200 iterations (before the
memory is breaking the process) and the output error is still the same
value.
The only way to have the error value "moving" is to set a decay, tried with
0.01 and error is getting lower, but very slowly ..

If I compare the error, on the same dataset with another library (Fann), the
error is very different:

Torch MSE error: about 0.18 after 100 iteration (and not decreasing), Fann
MSE error: about 0.047 after 100 iterations and still decreasing ...

Maybe I am not using Torch correctly, I have been playing with all samples
and tried different functions, transient layers etc .. but I am still
puzzled by the results.


Any help/guidance would be greatly appreciated.

Regards
Franck

[Torch5-devel] API Changes: MapStorage

[Ronan C. <ro...@co...>]

Hey

   (1) Due to some serious bugs in MapStorage, I had to modify the  
API. Sorry.
   MapStorage no longuer exists. It is now handled by a new  
constructor of Storage.

   Ex: torch.CharStorage(fileName, shared)

   If the optional argument "shared" is true, the mapping will be  
shared amongst all processes on the computer. In that case, the file  
must be with read-write access. Modifications made on the storage  
will then be reflected in the file. This corresponds to the old  
behavior of MapStorage.

   If "shared" is false (or not provided), the file must be at least  
readable. Modifications made on the storage will not affect the file.

   This fixes the bug reported by Gregoire about segfault when  
freeing a MapStorage.

   (2) A new shortcut to create a 1D Tensor over a Storage has been  
added.
         Ex: s = torch.DoubleStorage(100)
                t = torch.Tensor(s) -- in previous versions you had  
to do torch.Tensor(s, 1, s:size())

The doc has been updated accordingly for (1) and (2) in the SVN.

Cheers
Ronan.

Re: [Torch5-devel] torch5 svn

[Ronan C. <ro...@co...>]

Hey

   The SVN repository has been cleaned up. You will have to check it  
out again, I am sorry for the inconvenience. The instructions differ  
slightly than previous ones. It is now:

svn checkout https://torch5.svn.sourceforge.net/svnroot/torch5/trunk  
torch

(while before it was 'svn checkout https://torch5.svn.sourceforge.net/ 
svnroot/torch5/trunk/torch')

The second argument 'torch' is the name you want for the directory.   
You can call it whatever you want.

   A contrib packages SVN repository has been created. It is  
optional, but you can checkout it with:
cd torch
svn checkout https://torch5-contrib.svn.sourceforge.net/svnroot/ 
torch5-contrib/trunk contrib

Notes:
  (1) contrib packages should be checked out in torch main directory.  
They must be checked out in contrib in this main directory so that  
CMake can detect them.
  (2) to compile a contrib package, like "jpeg", you add - 
DWITH_CONTRIB_jpeg=1 when calling CMake. The package should be then  
compiled with the rest.
  (3) there is no trace of "external/" directory anymore in the  
torch5 SVN repository.

Ronan.

On Sep 16, 2008, at 4:46 PM, Ronan Collobert wrote:

> the svn repository is still under re-arrangement. i just submitted  
> the new clean dump svn file to sourceforge, and the sourceforge  
> guys should update it in the next hours. please, i *insist* do not  
> commit or even update in the meantime. i will keep you updated.
>
> ronan.
>

[Torch5-devel] luaT & MapStorage & SVN

[Ronan C. <ro...@co...>]

Hi,

*** LuaT

   I made major changes in luaT, such that it handles better class  
inheritance.

   From the Lua side, this has little impact.

   From the C side, few functions changed. In particular  
luaT_newmetatable() now takes
      * a string for the name of the class
      * a string for the parent class (or NULL) (this was previously  
an id)
      * a constructor (or NULL)
      * a destructor (or NULL)
      * a factory (or NULL)
   Once a constructor/destructor or factory is set, it cannot be  
changed, avoiding messing up from Lua. Also, the true metatable is  
hidden from Lua now, you can see only the metatable containing  
methods (so you cannot messup constructor or __gc from there).

   New feature: all functions of a class can now be accessed in the  
following way: torch.Tensor.fill(x, 1) instead of x:fill(1).

*** MapStorage

   As several people seemed interested in it, I added MapStorage in  
torch module. It works on the platforms we support.
   x = torch.DoubleMapStorage(filename) returns a kind of  
DoubleStorage which maps the contents of a file in memory.

*** SVN

   The website has been updated. Now Torch should be officially  
retrieved from SVN.

Ronan.

[Torch5-devel] luaT api changes

[Ronan C. <ro...@co...>]

hi

   i had to cleanup luaT before everybody starts to use it, to be a  
bit more consistent with luaL.
   here are some naming changes:

luaT_istorch became luaT_id
luaT_getmetatable became luaT_pushmetatable
luaT_typename became luaT_id2typename
luaT_id became luaT_typename2id
luaT_checkid became luaT_checktypename2id

new functions: luaT_typerror and luaT_typename

i changed the packages accordingly. should not bother you too much,  
as few of you were using luaT.

documentation coming soon.

ronan.

[Torch5-devel] moving to longs

[Ronan C. <ro...@co...>]

hey guys

   *** i made to major changes in torch 5.1
    - added a versioning system in object saved on disk
    - changed Tensor and Storage sizes from "int" to "long"

   we can now use all the gigs of ram we want :)

   the models you saved previously should be loadable with the new  
torch, thanks to the versioning system. if you have problems loading  
an old model, please let me know.

   *** there is one incompatibility coming out of this:
   if you were creating a tensor with something like
     sz = torch.IntStorage(2); sz[1] = 4; sz[2] = 5
     x = torch.Tensor(sz)

   this does not work anymore: you have to use LongStorage instead. i  
believe this was rare anyways, so it should not bother you too much.

cheers
ronan.

[Torch5-devel] map & apply bug

[Ronan C. <ro...@co...>]

Guys

   I corrected a bug occurring in map*() and apply() functions (torch  
& lab package).

This bug was appearing only if the function you provided as an  
argument returned sometimes a number, sometimes nothing (or nil). If  
it returned always a number or always nothing (or nil), the behavior  
was correct.

I also corrected the doc concerning these functions. In particular,  
lab.apply() was previously renamed as lab.map() to be consistent with  
torch package functions, but the doc was not updated (and wrong).

Cheers
Ronan.

[Torch5-devel] Lua QT

[Ronan C. <ro...@co...>]

Hey

[I]  A lot of stuff concerning QT has been added in the CVS.  
Everything has been written by Leon Bottou. This allow to make easy  
(and powerful) graphics/GUI directly from Lua. Several packages have  
been written for that purpose:

* qt: general utilities for interfacing Variants/Object/Classes from  
QT. Also functions for dealing with signals & slots.
* qttorch: able to convert tensors to/from QT images
* qtuiloader: can load QT GUI "ui" files (for e.g. written by the QT  
designer)
* qtwidget: interface to QT widgets

This packages are available if you run "qlua" (a variant of the "lua"  
program which knows about these packages) instead of "lua".

Another Torch package has been created:
* graphics: graphics drawing primitives interfaced to Cairo or QT. QT  
interface is used if QT is installed.

CMake detects if you have QT 4.3 or higher installed, in which case  
it compiles these modules.

Because of several annoying bugs in CMake, the minimum version  
required for CMake are now:
- 2.6.0 under Windows (native & Cygwin) [Cygwin now contains this  
version, and under native Windows platform you can also (re)-install  
the easy installer)
- 2.4.8 under MacOS X (update your MacPorts, or re-install with the  
easy Mac installer)
- 2.4.7 under Linux

The website has been updated accordingly, but more doc is coming soon.

Please, let me know if you experience any problem.

[II] Due to several complaints of people who did not installed  
fontconfig & freetype2 under Linux systems (and which got a  
compilation error because Cairo needs those to be compiled), the  
compilation of Cairo is not attempted under Linux if those packages  
are not installed.

Ronan.

[Torch5-devel] Torch 5.1 API changes

[Ronan C. <ro...@co...>]

Hey

   So I cleaned up the torch package in Torch 5.1.
   I updated the doc as well (in the CVS, not on the website yet).

   You might be interested in some functions like map, map2, apply  
which were undocumented until now :) I also improved the overloaded  
operators behavior.

Note: the order of arguments in map and map2 changed. I think it  
makes more sense like that. As probably nobody knew about them, it  
should be ok.

I will soon commit the new website.
Cheers,
Ronan.

Re: [Torch5-devel] torch5 newbie question

[Ronan C. <ro...@co...>]

> How do you decide whether a class should be a C++ class or a Lua class?

  If Lua will not slow it down much, do it in Lua. Not only it is faster 
to program, but it is also more powerful: classes in Lua can use Lua and 
C++ classes, while classes in C++ can only use C++ classes...

When you want to try quickly if an idea works, do it in Lua first as 
well... if it does not, no need to spend hours to debug your C++ code!

> Or why do you have for example in the nn package a 
> StochasticGradient.lua instead of a StochasticGradient.cpp?

StochasticGradient is basically a loop, which makes calls to functions. 
Not very CPU intensive! So Lua is great for that. It allows 
StochasticGradient to handle Lua GradientModules, in case you want to 
quickly (for trying some idea) design some dirty modules.

You can also provide it some Lua datasets, which is extremely handy: as 
long as you have a Lua class which handles :size() and [i] 
StochasticGradient will handle it as a dataset!

> And in the lab package you have a matelem.cpp AND a matelem.lua!?

All lab functions use intensively numbers and thus are coded in C++. 
However, functions like rand() zeros() ... can take both integers like 
rand(5,2) or a IntStorage as an argument. With integers they have to 
handle variable numbers of arguments, and that is much easier to handle 
in Lua than in C++ (as low cost). So this top part is done in Lua, which 
then calls the underlying C++ function: once again, the CPU intensive 
part (well, the calculations, numerical stuff) is done in C++, and the 
rest in Lua.

Ronan.

[Torch5-devel] torch5 newbie question

[Mikaela K. <mik...@gm...>]

Hi,
I suppose this is (will be) a very common question... Lets have it
documented.
How do you decide whether a class should be a C++ class or a Lua class?
Or why do you have for example in the nn package a 
StochasticGradient.lua instead of a StochasticGradient.cpp? And in the 
lab package you have a matelem.cpp AND a matelem.lua!?
Thanks,
Mika

[Torch5-devel] Concat in nn package

[Ronan C. <ro...@co...>]

Hey

   Jason pointed me out that Concat in nn package was not working. In 
fact, it was really buggy -- the implementation was not even finished. I 
corrected it, and improved it.

l = nn.Concat([dim]) takes now an optional dim argument: the dimension 
along which you want to concatenate (default 1) [that is new, before it 
was 1 with no option].

each modules added in Concat (with l:add(...)) must take the same number 
of inputs. their outputs are concatenated (along dim given in the 
constructor) and returned by the Concat layer.

you can feed tensors of any dimension to Concat (provided that the 
modules inside understand them). the outputed dimensions of each module 
might be different for the dimension "dim" but must (of course!) be the 
same in other dimensions.

ronan.

[Torch5-devel] new functions

[Ronan C. <ro...@co...>]

hey

   i added:

     - apply() in Tensor. x:apply(func) applies func to all element of 
x, and returns x. a bit like lab.apply() but no memory creation.

     - lab.range(xmin, xmax [, step]) returns a Tensor with values 
between xmin (included) and xmax (possibly included, depending on the 
step). steps between each value is 1 (default) or step (if given).

    - lab.randperm(n)  as in matlab, returns a Tensor containing a 
permutations of the integers 1..n. Similar to random.shuffledIndices() 
(but shuffledIndices() returns a Storage)

ronan.

[Torch5-devel] apply() and torch5-devel mailing list

[Ronan C. <ro...@co...>]

Hey

---- disclaimer

i created a torch5-devel list. so if you want to discuss something on 
the development of torch, write here. new added features should also be 
advertised here. i will not send any mail personally anymore, so i 
strongly encourage you to subscribe to the list if you are interested by 
this kind of news.

https://lists.sourceforge.net/lists/listinfo/torch5-devel

---- end of disclaimer

   I added in lab package the lab.apply() function.

Suppose you have a 1000x10000 random tensor and you want to apply a tanh 
on it. You can:

x = lab.rand(1000,10000)

-- superfast, takes 0.7s
z = lab.tanh(x)

-- superslow, takes 44.7s
z = torch.Tensor():resizeAs(x)
for i=1,x:size(1) do
    for j=1,x:size(2) do
       z[i][j] = math.tanh(x[i][j])
    end
end

-- superslow, takes 15.5s
-- we can do it here, because we know the tensor is contiguous in memory
z = torch.Tensor():resizeAs(x)
local xs = x:storage()
local zs = z:storage()
for i=1,xs:size() do
    zs[i] = math.tanh(xs[i])
end

-- supercool, takes 1.4s
z = lab.apply(x, math.tanh)


So this lab.apply() function is extremely efficient (considering it does 
lua calls underneath) and should be used instead of loops whenever you 
can. (well, except of course if an optimized C function already exists 
like for tanh())

Ronan.