Sander Dieleman

Here's another approach for tackling this "optimal layout" issue in frameworks using the computational graph paradigm (such as Theano and TensorFlow): stick with the canonical NCHW or NHWC layout on...

I can try coding up those first 3 probably, but GoogLeNet in Theano is still a challenge at the moment, and it would require some tricks that would slow it...

Awesome! Very happy the FFT approach is still being pursued and it's starting to bear fruit :) It's also cool that it's using CuFFT and doesn't require a custom implementation,...

Cool, would love to see a preprint of that at some point, if you and your colleagues are willing to share :)

Awesome!

I hadn't noticed, thanks for pointing that out! I don't know how to build AlexNet in Theano either, it's not easy to do 'same' convolutions across the different convolution implementations...

Ok, then the problem is the 3x3 pooling with stride 2. Someone is working on implementing strided pooling, but it hasn't been merged: https://github.com/Theano/Theano/issues/2196 It's possible to do a custom...

Hey, don't look at me, I've brought this up in the past :) (See https://github.com/Theano/Theano/issues/2118 and https://github.com/Theano/Theano/issues/2196 ) Yeah, the cudnn pooling has been wrapped, but ideally you don't want...

Sure, but I'm thinking of stuff like the custom output layer I used for the Galaxy Challenge, to incorporate the constraints on the predicted probabilities (this monster: https://github.com/benanne/kaggle-galaxies/blob/master/custom.py#L576 ). There...

Merry Christmas and a happy new year to everyone as well :) @nouiz: To be honest, I don't really care about CPU support for every feature either. All the machines...