Hi guys, So I started implementing a full example using Robb's description of U-Net as well as the Dental Challenge dataset they competed against. You can see my ipython notebook here <http://nbviewer.jupyter.org/gist/mongoose54/d88da59c24d452bef002436156051fba>. The issue I am facing is that my predicted images are grayscale (e.g. having normalized pixel values 0-1). I thought that using the mean or argmax I would get the predicted masks. Unfortunately as you can in the notebook the histogram of the predicted images is not bi-modal. I am wondering if I am using the wrong loss function (i.e. categorical cross entropy) to optimize on and I should use some other function. Let me know of your thoughts.
Thanks, Alex On Tuesday, July 19, 2016 at 1:27:06 PM UTC-7, [email protected] wrote: > > Yes, I'll install Ubuntu and try your code under Linux. The only library I > found to stably work under Windows is Matlab parallel computing toolbox: > there is an example of a classical CNN but I guess it is not flexible > enough to build something similar to u-net. Also I do not know how > efficient it is and would prefer to start with a library with large > community. > -- --- 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.
