You're running out of memory because your model is enourmous, particularly w4. w4 by itself looks to be around 100 GB, if you're using 32-bit floats. If you want to run on any kind of normal hardware you need to reduce the number of parameters.
You could use a smaller fully connected network, say with no set of weights greater than 100x100, but for images convolutional networks with maxpooling or other types of downsampling are very efficient, so they scale up very well. On Tuesday, August 16, 2016 at 3:46:11 PM UTC-4, Станислав Кусков wrote: > > how about "uot of memory" error? I have it for traning on GPU. On CPU i > have problem with RAM memory. > > i can solve it on programm, or changing hardware only? > > вторник, 16 августа 2016 г., 22:36:03 UTC+3 пользователь Robb Brown > написал: >> >> Convolutional networks can be used to efficiently work with large images. >> We typically work with volumes that are 200x200x200 or so. >> >> Another approach would be to use a fully connected network with lower >> capacity. That will often work better anyway because it will avoid some >> over fitting. >> > -- --- 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.
