[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