Support for 1D and Non-square convolutional kernels - Githubissues

hughperkins / DeepCL

OpenCL library to train deep convolutional neural networks

Mozilla Public License 2.0

867 stars 199 forks source link

Support for 1D and Non-square convolutional kernels #37

Open brazilla opened 8 years ago

brazilla commented 8 years ago

Currently only 2D square convolutional kernels are supported. Thx.

hughperkins commented 8 years ago

Understood. Involves a fair amount of work. Will ponder. Anyone else who arrives in this thread, and wants non-square kernels, please add a +1 into this thread, to motivate me a bit :-)

hughperkins commented 8 years ago

Oh, hmmm, kernels.... is this for Inception?

brazilla commented 8 years ago

Mainly for NLP and other sequence processing.

hughperkins commented 8 years ago

Ah, hmmm, right.

Presumably you'd want the input 'images' to also be non-squared too? Presumably, quite very not-square, like 1-dimensional?

brazilla commented 8 years ago

Yeah,exactly.

hughperkins commented 8 years ago

Hmmm... does sound potentially quite useful generally. By the way, just out of curiosity do you know any papers that use this technique?

brazilla commented 8 years ago

http://nal.co/papers/Kalchbrenner_DCNN_ACL14 http://nlp.stanford.edu/~socherr/EMNLP2013_RNTN.pdf http://www-nlp.stanford.edu/pubs/SocherLinNgManning_ICML2011.pdf

hughperkins commented 8 years ago

Thanks!

wbernoudy commented 8 years ago

+1. Has there been any progress with this?

hughperkins commented 8 years ago

Hmmmm .... well.... clnn contains a wrapper for 1-d convolutions. https://github.com/hughperkins/clnn/blob/master/TemporalConvolution2.lua I confess I'm unlikely to dedicate any time to implementing this in DeepCL in the short-term. I dont imagine it's tons of work, as long as you:

make it only for 1-d, ie not generally rectangular support
use the existing im2col convolution implementation. It's fast, and means you wont need to roll up your sleeves and dive into opencl too much (might need to ... a bit... for im2col ... maybe :-P) It is however non-trivial enough that I'm not just going to spend a few hours one morning, and make it work. It's probably about a weeks work, I would think, ballpark. 40-80 man-hours or so.

hughperkins commented 8 years ago

Note: I think an appropriate way forward for this might be to migrate to use https://github.com/BVLC/caffe/issues/4155 , so that any kernels are shared across OpenCL projects.

dex1990 commented 7 years ago

hi,@hughperkins did DeepCL support 3D square convolutional layers ? e.g. 3x3x64, which means size of filter is width 3,height 3 and depth 64.

hughperkins commented 7 years ago

Correct. Can you confirm what makes you feel this might not be the case?

On 15 November 2016 10:35:01 GMT+00:00, dex1990 notifications@github.com wrote:

hi,@hughperkins did DeepCL support 3D square convolutional layers ? e.g. 3x3x64, which means size of filter is width 3,height 3 and depth 64.

You are receiving this because you were mentioned. Reply to this email directly or view it on GitHub: https://github.com/hughperkins/DeepCL/issues/37#issuecomment-260606541

Sent from my Android device with K-9 Mail. Please excuse my brevity.

dex1990 commented 7 years ago

Thank you for your answer. I have not found any example on it, maybe i missed some relevant examples. I have trained model from this project: https://github.com/huangzehao/caffe-vdsr some layers of this network have input size: 40x40x64 and 64 filters with size: 3x3x64, so the output size is 40x40x64. I want to test the trained network with DeepCL.

0StackOverflow0 commented 7 years ago

I'm interested in adding 3D convolutional filters (once I do some more reading), but I'm hoping to have a proposal for stride in the not too distant future.

I wonder how hard 1D would be after doing that.

hughperkins commented 7 years ago

I'm hoping to have a proposal for stride in the not too distant future.

Sounds good :)

I wonder how hard 1D would be after doing that.

Its not conceptually hard, but it might involve a certain amount of legwork :)