These are chat archives for torch/torch7

14th
Jun 2015
Hokie23
@Hokie23
Jun 14 2015 00:07
@culurciello well, would some sort of contrast jitter or color jitter of the background help? Too late to collect more data.
John Armstrong
@jarmstrong2
Jun 14 2015 00:08
help please...
Laurens van der Maaten
@lvdmaaten
Jun 14 2015 00:10
@jarmstrong2 Cutorch is an implementation of the core torch library for Cuda. Cunn is the Cuda implementation of the nn neural network package. So it depends on what you want to do. In many usecases, I suppose you would actually be using both cutorch and cunn.
John Armstrong
@jarmstrong2
Jun 14 2015 00:12
@lvdmaaten thanks! Anything you would recommend for speeding up training of rnns as I mention above, I would appreciate any input. Thanks again.
Laurens van der Maaten
@lvdmaaten
Jun 14 2015 00:16
Hard to say... In general, you want to minimize the amount of copying that goes on between the host and the GPU, and you want to minimize the amount of memory allocation that is going on on the GPU. So try and allocate memory as few times as possible (instead, better re-use pre-allocated memory). But bear in mind that large networks may take a long time to train: some of the biggest convnets even need several weeks of training time.
Hokie23
@Hokie23
Jun 14 2015 00:21
@culurciello also there are plenty of lighting changes and shadows as people are driving. So it's not just 2 backgrounds
Hokie23
@Hokie23
Jun 14 2015 01:18
@lvdmaaten Hey! So with the weight decay of 1e-03 both the train and test accuracies for the first epoch are around 20. And they predicted only the first class in both cases. I am going to wait for another 2 or 3. At least the test and train errors are consistent.
Laurens van der Maaten
@lvdmaaten
Jun 14 2015 01:20
Okay well at least you're not overfitting anymore ;)
Hokie23
@Hokie23
Jun 14 2015 01:20
Haha yeah. Thank you!
Laurens van der Maaten
@lvdmaaten
Jun 14 2015 01:20
You may have to tune things like weight decay and dropout rate very carefully (on a held-out validation set).
Hokie23
@Hokie23
Jun 14 2015 01:22
I have the drop out in a few of the conv layers as well as all the fully connected layers . Both at .5, will tune them once I get to a nice point.
Soumith Chintala
@soumith
Jun 14 2015 02:53
@lvdmaaten :clap: :clap: :clap: :clap: :clap: :clap: :clap: :clap: :clap:
Hokie23
@Hokie23
Jun 14 2015 03:00
Hey @soumith. Find any interesting papers at this CVPR?
Soumith Chintala
@soumith
Jun 14 2015 03:14
@Hokie23 "recurrent convolutional neural network for object recognition". It was something we were planning to do, but the dude did it first, and it works pretty well apparently. super-easy to implement as well.
http://www.cv-foundation.org/openaccess/content_cvpr_2015/html/Liang_Recurrent_Convolutional_Neural_2015_CVPR_paper.html
Hokie23
@Hokie23
Jun 14 2015 10:48
@soumith thanks! Take that CRFs!
John Armstrong
@jarmstrong2
Jun 14 2015 11:16
Is there anyway to change a model back to CPU after it has been adapted to GPU, for instance say you have model = nn.Linear(1,1):cuda(), can I switch it back to its non-cuda state?
xuwang
@xuwangyin
Jun 14 2015 11:43
John Armstrong
@jarmstrong2
Jun 14 2015 11:56
@xuwangyin bless you!
Alfredo Canziani
@Atcold
Jun 14 2015 21:06
@eladhoffer (@culuriciello, @jhjin), I have 12 ms and 3 ms of random loading time on HDD and SSD for images of 3x128x128. JPEG decoding time is constant to 8 ms per sample though. I'm not sure how you achieve 3 ms for sequential reading and decompression. I'm currently using the JPEG-turbo8 libraries and ImageMagicks for doing this.
So, I'm constantly above 11 ms per sample.
Elad Hoffer
@eladhoffer
Jun 14 2015 21:46
Are you loading from raw JPEG? My loading times are from pre-archived LMDBs. Decoding is done to byte tensor btw