Couldn't load models/VGG_ILSVRC_19_layers.caffemodel

[hangbomb]

Hi,

I installed neural-style and have been trying to run th neural_style.lua -gpu -1 -print_iter -1 on Ubuntu 14.04, but I'm keep getting following error.

Couldn't load models/VGG_ILSVRC_19_layers.caffemodel
/home/XXX/torch/install/bin/luajit: neural_style.lua:59: attempt to index a nil value
stack traceback:
    neural_style.lua:59: in function 'main'
    neural_style.lua:437: in main chunk
    [C]: in function 'dofile'
    XXX/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:131: in main chunk
    [C]: at 0x00406670

Is there any solution?

[FlorianWendelborn]

Did you download the models?

[jcjohnson]

This isn't causing your current problem, but you should also use -print_iter 1 instead of -print_iter -1.

[hangbomb]

Thank you for your advice.

Yes, I had downloaded the models before running neural style. Also I tried -print_iter 1, but still getting this.

Couldn't load models/VGG_ILSVRC_19_layers.caffemodel
/home/XXX/torch/install/bin/luajit: neural_style.lua:59: attempt to index a nil value
stack traceback:
    neural_style.lua:59: in function 'main'
    neural_style.lua:437: in main chunk
    [C]: in function 'dofile'
    XXX/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:131: in main chunk
    [C]: at 0x00406670

Is there any possibility that my laptop is the problem?

[spinda]

I also have this problem. I've downloaded all the model files, and I'm also running on a laptop (with -gpu -1).

Couldn't load models/VGG_ILSVRC_19_layers.caffemodel
/home/spinda/workbench/torch/install/bin/luajit: neural_style.lua:60: attempt to index a nil value
stack traceback:
    neural_style.lua:60: in function 'main'
    neural_style.lua:441: in main chunk
    [C]: in function 'dofile'
    ...ench/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:131: in main chunk
    [C]: at 0x00405ea0

[spinda]

It seems that loadcaffe_wrap.load(params.proto_file, params.model_file, params.backend) returns nil for some reason.

[hughperkins]

That error message comes from the depths of loadcaffe library. loadcaffe_wrap calls it in lin 17:

  C.loadBinary(handle, prototxt_name, binary_name)

The error message comes from the end of the loadBinary method, in loadcaffe.cpp, inside loadcaffe library:

void loadBinary(void** handle, const char* prototxt_name, const char* binary_name)
{
  caffe::NetParameter* netparam = new caffe::NetParameter();
  ReadProtoFromTextFile(prototxt_name, netparam);
  bool success = ReadProtoFromBinaryFile(binary_name, netparam);
  if(success)
  {
    std::cout << "Successfully loaded " << binary_name << std::endl;
    handle[1] = netparam;
  }
  else
    std::cout << "Couldn't load " << binary_name << std::endl;
}

You probably want to:

• make a tiny test case that just uses loadcaffe (no loadcaffe_wrapper etc)
• raise the issue on loadcaffe repo issue tracker

[hughperkins]

(note that personally I'm using hte normalised files. the md5sums are:

$ md5sum models/*
ccbbdda59210208be39f8974f5b5765e  models/VGG_ILSVRC_19_layers_deploy.prototxt
6adcfbc93e8f6762e6421515940526f4  models/vgg_normalised.caffemodel

)

[spinda]

I can verify that these md5sums match the model files I'm using.

[hughperkins]

Ok. i'm on ubuntu 14.04 too. My loadcaffe gitlog is:

$ git log -n 10 --oneline
63fe14e Merge pull request #24 from chirag64/patch-1
a1781e2 Removed unnecessary comma character
6632354 make inn and cunn deps optional
bb2469c Merge pull request #21 from fmassa/pool_pad_fix
a30c883 Fix pooling when pad != 0
51a2ff8 more verbose print in case of unknown module
4762b38 fixed a bug with ceil average max pooling
5b17397 Update README.md
e39087f Merge pull request #17 from szagoruyko/cpu
7aff9ec cpu support

(looks like I'm using a version from around 8th september)

[spinda]

I was using loadcaffe straight from luarocks install loadcaffe. Removing that and switching to 63fe14e from git unfortunately didn't make the problem go away.

A minimal test case producing the issue is:

require 'loadcaffe'
loadcaffe.load('models/VGG_ILSVRC_19_layers_deploy.prototxt','models/VGG_ILSVRC_19_layers.caffemodel','nn-cpu')

[hughperkins]

Hmmm, you know what, iget the same error :-P And yet... I can use neural-style... perahps there is some gap in my analysis above...

[spinda]

That didn't make any difference, unfortunately. Still getting the same error.

[hughperkins]

Works ok for me:

require 'loadcaffe'
loadcaffe.load('models/VGG_ILSVRC_19_layers_deploy.prototxt','models/vgg_normalised.caffemodel','nn')

[hughperkins]

This works for me too:

local ffi = require 'ffi'
require 'loadcaffe'
local C = loadcaffe.C

local backend = 'nn-cpu'
local prototxt_name = 'models/VGG_ILSVRC_19_layers_deploy.prototxt'
local binary_name = 'models/vgg_normalised.caffemodel'
local handle = ffi.new('void*[1]')

-- loads caffe model in memory and keeps handle to it in ffi
local old_val = handle[1]
C.loadBinary(handle, prototxt_name, binary_name)

[spinda]

Ah ha! I was using the default files, rather than the normali(z|s)ed ones. Running neural-style with the custom file options seems to have started things working for me:

th neural_style.lua -gpu -1 -proto_file models/VGG_ILSVRC_19_layers_deploy.prototxt -model_file models/vgg_normalised.caffemodel -style_image style.jpg -content_image content.jpg

Thanks!

[hughperkins]

Cool :-)

[spinda]

As it turns out, my copy of the model file was corrupted after all (see szagoruyko/loadcaffe#31). Thanks for all the help.

[jcjohnson]

Good catch! Glad you figured it out.

[hughperkins]

For ease of access, note that the correct md5sum is:

b5c644beabd7cf06bdd9065cfd674c97 VGG_ILSVRC_19_layers.caffemodel

[subzerofun]

I just had the exact same issue – a corrupt .caffemodel file. Just found this page after i've already uninstalled/reinstalled torch and all the extra packages i've accumulated over the time... Google always leads you to the right results :-)