Caffe's bvlc_googlenet and Lasange's GoogLeNet do not produce the exact same results about recipes

[f0k]

How much do they differ? Is this on CPU or GPU? Do you use cuDNN?

[yassersouri]

I use GPU and cuDNN for both caffe and lasagne. Also I have just installed Theano and Lasagne from master today.

The l2 norm of their difference for an example image is 0.14590697. (the example image the is cat image used in the caffe tutorial)

plt.plot(lasagne_out - caffe_out)

Here is a plot of the difference, as I said they are very similar.

[yassersouri]

Well this was not a good example (since some might think the preprocessing is not the same for caffe and lasagne).

For the input np.zeros((1, 3, 224, 224), dtype=np.float32) the l2 norm of the difference is 0.096159138.

Also here is the plot for plt.plot(lasagne_out - caffe_out).

[ebenolson]

Hi @yassersouri, thanks for your report.

It looks like several of the pooling layers should use ignore_border=False but are not. My guess is that when I wrote this I was using an older version of Lasagne (before Lasagne/Lasagne#374).

With this change I get an l2 norm of the difference for zero input of 9.7e-8, I will push an updated version shortly.

[yassersouri]

@ebenolson Thanks.

[f0k]

Oh, good catch! Sadly, this also means it won't be able to use cuDNN for pooling any more (it always ignores the border).

[yassersouri]

Do other models in the model zoo need similar updates?

[ebenolson]

I don't think so - I seem to remember this being a peculiarity of googlenet (I think this is why I was using both Pool2DLayer and Pool2DLayerDNN).

However, it would be good to verify that, and add some numerical tests to catch any future problems. PR's welcome 😄