These are chat archives for torch/torch7

4th
Sep 2015
AAnoosheh
@AAnoosheh
Sep 04 2015 00:42 UTC
I'm new to Torch. What exactly is the accGradParameters function supposed to do? It implies "Accumulate Gradient w.r.t. Parameters" but I still don't know what the word "Parameters" means exactly, either.
Justin
@jcjohnson
Sep 04 2015 04:49 UTC
@soumith Thanks for posting your benchmark result! Now I know for sure that something is wrong with my setup. I was indeed on old drivers, but even after upgrading to 352.41 I'm still getting the same horrible performance on cudnn convolutions ... any other idea what might be wrong?
Soumith Chintala
@soumith
Sep 04 2015 05:04 UTC
@jcjohnson not really sure what else to suspect. CUDA 7.5?
James Bradbury
@jekbradbury
Sep 04 2015 06:27 UTC
has anyone seen /usr/local/bin/luajit: C++ exception without a traceback at a totally arbitrary (but reproducible) point during training of a torch model? i'm about to start what's likely to be a rather torturous process of figuring out what exactly triggers the error...
Francisco Massa
@fmassa
Sep 04 2015 07:06 UTC
@vislab2013 @soumith I'm aware of this problem with SpatialPyramidPooling. I pushed a PR to nn torch/nn#284 to fix it, but it didn't get accepted.
Using a :reshape instead of :view is not that easy, as we would require in the constructor the number of input channels (Reshape doesn't accept -1 indices).
An easy solution is to add a nn.Copy(nil,nil,true) just after the nn.View in line https://github.com/szagoruyko/imagine-nn/blob/master/SpatialPyramidPooling.lua#L13 , but that makes an unnecessary memory copy in the forward pass...
Eduard Trulls
@etrulls
Sep 04 2015 08:36 UTC
is there a way to not have itorch.image automatically scale the image range to [0,1]?
vislab2013
@vislab2013
Sep 04 2015 09:54 UTC
@fmassa tyvm! Unnecessary copies always makes me frown but, for the time being, it's a quick fix and i'll take it.
Cheers
Francisco Massa
@fmassa
Sep 04 2015 10:03 UTC
@soumith do you think a module called nn.Contiguous could be useful in situations like that ?
Eugenio Culurciello
@culurciello
Sep 04 2015 10:50 UTC
Hi guys,what do you use to check image types and detect image types that are not supported? Do you use the "image" package or imagemagick?
vislab2013
@vislab2013
Sep 04 2015 14:03 UTC
is there any module for torch that may work as a gate inside a network? For example, block a certain branch if conditions X, Y, Z aren't met by the input.
Francisco Massa
@fmassa
Sep 04 2015 14:11 UTC
@vislab2013 it's not exactly what you asked, but it's a bit related https://github.com/torch/nn/blob/master/doc/table.md#nn.MixtureTable
vislab2013
@vislab2013
Sep 04 2015 14:18 UTC
@fmassa thanks, i'll take a look at it.
Soumith Chintala
@soumith
Sep 04 2015 14:58 UTC
@fmassa nn.Contiguous sounds pretty useful. I was thinking the same.
Eugenio Culurciello
@culurciello
Sep 04 2015 15:04 UTC
Dear @soumith in https://github.com/soumith/imagenet-multiGPU.torch is there a way to select WHICH gpus to use?
Eugenio Culurciello
@culurciello
Sep 04 2015 15:13 UTC
@soumith in order to use GPUs 3,4 in a system with 4 GPUS, for examples, we would need to change https://github.com/soumith/imagenet-multiGPU.torch/blob/master/models/alexnet_ccn2.lua#L38 to "for i=opt.GPU,nGPU do", right?
Francisco Massa
@fmassa
Sep 04 2015 15:14 UTC
@culurciello just use CUDA_VISIBLE_DEVICES=2,3 before launching your script
Eugenio Culurciello
@culurciello
Sep 04 2015 15:14 UTC
In production web api system, do you guys use imagemagicks for loading and preparing images?
@fmassa thank you for the tip, great!
Francisco Massa
@fmassa
Sep 04 2015 15:34 UTC
@soumith cool, I'll implement it then
duy huynh
@duyhtq
Sep 04 2015 20:46 UTC
@culurciello are you running torch in production web api system? would love to learn more how you do it. we're trying to (1) find a similar framework to flask (for python) for torch and (2) figure out how to persist / cache the torch model on the server side so we don't have to reload it for every request. any advice is much appreciated.