We could wrap the cuda code in Theano. We don't have time now to do it, but we can guide someone that would be interrested to do it.
A first starting point is the doc about how to make new Ops in Theano: http://deeplearning.net/software/theano/tutorial/extending_theano.html Fred On Thu, Aug 25, 2016 at 11:04 PM, Arjun Jain <[email protected]> wrote: > Thanks a lot for sharing the video! :) > > On Thu, Aug 25, 2016 at 10:18 PM, 7VoltCrayon <[email protected]> > wrote: > >> Yes, the slides are a bit difficult to understand on their own, I haven't >> looked at what Caffe is doing with group yet, from the discussion it looks >> like something similar but I could be wrong (and they don't mention the 1 x >> 1 convolutions to mix the resulting feature maps). Until I take a closer >> look at it, this lecture by the author has a ~5 minute section where he >> explains the separable convolution slides succinctly: https://youtu.be/V >> hLe-u0M1a8?t=1087 >> >> On Wednesday, 24 August 2016 20:45:22 UTC-7, Arjun Jain wrote: >>> >>> Hi 7VoltCrayon, >>> Thanks for pointing to this grouped convolution. I am not sure if I >>> understood the slide 100%, but to me it sounds similar to caffe's "group >>> by" (https://github.com/BVLC/caffe/issues/778), is that true? If not, >>> can you help me spot the differences? >>> >>> Thanks a lot. >>> >>> On Thu, Aug 25, 2016 at 3:29 AM, 7VoltCrayon <[email protected]> wrote: >>> >>>> Thank you for the reply. I see that in TensorFlow, this is implemented >>>> at the CUDA level >>>> <https://github.com/tensorflow/tensorflow/blob/d42facc3cc9611f0c9722c81551a7404a0bd3f6b/tensorflow/core/kernels/depthwise_conv_op_gpu.cu.cc> >>>> (linked), >>>> if implementing this in Theano, would it be possible to get a fast >>>> implementation using pre-existing Theano ops? Or is this something that >>>> needs to be done at a C++ / CUDA level? >>>> >>>> On Wednesday, 24 August 2016 13:48:59 UTC-7, nouiz wrote: >>>>> >>>>> no, but if someone is interrested, it can be done in Theano too. >>>>> >>>>> Fred >>>>> >>>>> On Wed, Aug 24, 2016 at 4:09 PM, 7VoltCrayon <[email protected]> >>>>> wrote: >>>>> >>>>>> Does theano have the equivalent of TensorFlow's separable_conv2d >>>>>> function? >>>>>> Where it implements a separable factorized convolution as described in >>>>>> these slides: http://vincent.vanhoucke.com/ >>>>>> publications/vanhoucke-iclr14.pdf?attredirects=0 >>>>>> >>>>>> -- >>>>>> >>>>>> --- >>>>>> You received this message because you are subscribed to the Google >>>>>> Groups "theano-users" group. >>>>>> To unsubscribe from this group and stop receiving emails from it, >>>>>> send an email to [email protected]. >>>>>> For more options, visit https://groups.google.com/d/optout. >>>>>> >>>>> >>>>> -- >>>> >>>> --- >>>> You received this message because you are subscribed to the Google >>>> Groups "theano-users" group. >>>> To unsubscribe from this group and stop receiving emails from it, send >>>> an email to [email protected]. >>>> For more options, visit https://groups.google.com/d/optout. >>>> >>> >>> -- >> >> --- >> You received this message because you are subscribed to the Google Groups >> "theano-users" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to [email protected]. >> For more options, visit https://groups.google.com/d/optout. >> > > -- > > --- > You received this message because you are subscribed to the Google Groups > "theano-users" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > For more options, visit https://groups.google.com/d/optout. > -- --- You received this message because you are subscribed to the Google Groups "theano-users" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
