Adapted torch's new API to fix #179

[petronny]

#cnn will return 0 instead the number of layers now.

[htoyryla]

When was this change made in loadcaffe? I just wonder is it possible that the problem in #179 , posted in March, could have been caused by this. I last installed loadcaffe as recently as August and did not have this problem then. I see no major commits to loadcaffe recently. I just wonder if there is some other reason causing #cnn to fail (I think I have seen it once or twice a long time ago, not in neural-style though).

[petronny]

Oh, it should be the changes of torch, sorry.
BTW, all torch stuff on my system is latest.
The usage of for i,layer in ipairs(cnn.modules) can be found in:
https://github.com/torch/demos/blob/master/person-detector/model.lua

[htoyryla]

I don't immediately see why torch would break it either. Made this minimal test script, which uses loadcaffe to load vgg19 into cnn and gets the dimension of both cnn and cnn.modules. Both work on both of my machines, the second at least has quite recent torch.

require 'loadcaffe'
cnn = loadcaffe.load("models/VGG_ILSVRC_19_layers_deploy.prototxt", "models/VGG_ILSVRC_19_layers.caffemodel", "nn"):float()
print(#cnn)
c = cnn.modules
print(#c)

[petronny]

I get

0
46

here.

[htoyryla]

OK, it appears that your method is safer. Just wonder what could cause the difference. Especially if this happened already in March for someone, so the real reason why #cnn fails is a mystery. Perhaps torch was built to use a different lua interpreter (my th uses luajit).

[htoyryla]

Installed torch again by luarocks, still getting 46 and 46.

[petronny]

My torch is installed from Archlinux user repository.
Maybe I should take a look at the PKGBUILD.

[ptekchand]

I ran into the same issue and commented on the line:

for i = 1, #cnn do 

f86babf#commitcomment-23341533
changing it to

for i = 1, #cnn.models do

worked.
If I understand correctly, one might run into this issue because a lua 5.2 compatibility/feature requirement (which wasn't enabled in luajit which was built on my installation - Windows 10 x64 MSVC 2015 x64 with the cunn backed).