On Fri, Aug 26, 2016 at 4:56 PM, Dustin Ezell <[email protected]> wrote:
> I'm curious if learning from features in the complex domain is better than >> learning from magnitude spectra or other common compromises made when using >> complex valued training data. I read somewhere once that there was a >> noticeable difference in models trained using a 2 real numbers vs actually >> using a complex number. The literature surrounding complex valued ANNs >> is... sparse. > > > > >> I'm not familiar with the C code portions of Theano. Where would I start >> looking to get familiar with this? I'm guessing at present there's no code >> to represent a complex matrix in a GPUArray? Or if there is basic things >> like broadcasting, multiplying, etc will need to be implemented first. >> Gradients would then follow. Then complex valued models would be possible... > > The complex in libgpuarray isn't working. If you want to work on that, open an issue there. It will get more seen by its main develeoper. I don't remember what isn't working. Part of the problem is that there is no C complex number like float. Fred > > On Friday, July 8, 2016 at 3:16:02 PM UTC-5, nouiz wrote: >> >> There is some work at the C code level. Are you familiar with it? >> >> Also, in the time you told, this will just be enable the complex number >> execution. with that available, you won't be able t make the grad on >> complex number work. >> >> Maybe the first step is to make the graph work on CPU with your complex >> number, we don't cover all operation there. >> >> Also, an important point to evaluate the time it will take is which >> operation you need on complex number. If it is just fft, we have a work >> around that represent the complex as an extra dimensions of size 2. (and >> this is already working). >> >> So which operation do you want in complex number? >> >> Fred >> >> On Fri, Jul 8, 2016 at 2:44 PM, Dustin Ezell <[email protected]> wrote: >> >>> How much work is left on that part? I could throw a few weeks of my >>> spare time at it if it's a task that's likely tractable for a newcomer. >>> >>> On Wednesday, July 6, 2016 at 10:00:40 AM UTC-5, nouiz wrote: >>>> >>>> We still don't have a good complex number support in Theano. The new >>>> GPU back-end could support it on the GPU, but we didn't finish that part as >>>> we don't need it. If you want that do be finished, we will need help for >>>> this. >>>> >>>> Frédéric >>>> >>>> On Tue, Jul 5, 2016 at 1:48 PM, Dustin Ezell <[email protected]> >>>> wrote: >>>> >>>>> What is the state of complex number support on GPU in Theano? They >>>>> were officially unsupported as of 6.0 but that was 3 years ago. >>>>> >>>>> -- >>>>> >>>>> --- >>>>> 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.
