scott89/fcnt

fcnt's Issues

caffe install

When I install caffe you provided, I got some troubles. when I executed the order "make runtest" . I got some errors as follows
F0401 04:20:30.789280 147288 conv_layer.cpp:54] Check failed: group_ == 1 (3 vs. 1)

The environments is as follows:
OS: Centos 7.0 64bit
cuda 7.5
cudnn4.0

Thanks

cudnn version

Could you tell me the version of your cudnn? When I compile your caffe-fcnt, i meet the following questions:
./include/caffe/util/cudnn.hpp(59): error: identifier "cudnnTensor4dDescriptor_t" is undefined
./include/caffe/util/cudnn.hpp(64): error: identifier "cudnnTensor4dDescriptor_t" is undefined
./include/caffe/util/cudnn.hpp(72): error: identifier "cudnnTensor4dDescriptor_t" is undefined
./include/caffe/util/cudnn.hpp(97): error: identifier "cudnnTensor4dDescriptor_t" is undefined
4 errors detected in the compilation of "/tmp/tmpxft_00004ef8_00000000-16_math_functions.compute_50.cpp1.ii".
make: *** [.build_release/src/caffe/util/math_functions.cuo] error 1

I think this is caused by the dismatch of cudnn version, but I tried to use cudnn 2.0, 3.0, 4.0, 5.0. All of these do not work, my cuda version is 7.5. Any help is appreciated and looking forward to your early reply.

train prototext file

Which prototxt file did you use for training? In feature_net.prototxt I can't find a reference to the dataset and output layer.

can't download paper

we can't download your paper on your paper link, could you check it please?

How to set the seq param for run.m?

I set the seq = VideoReader( ) , but I cannot add seq.init_rect attr , How to complete it ?

Where in the code is g_i and h_ii computed?

Hey
Thanks for sharing your code. I really like your paper. However I still cannot understand parts of the paper, so I tried looking into the code but couldn't find the solution there either. So my question is: Where and how do you compute the numbers g_i and h_ii which are used to compute the significance according to the paper?

It says si = - g_i f_i + 1/2 h_ii f_i^2

I guess this part calls the sel-CNN

select_snet_solver_def_file = 'solver/select_snet_solver.prototxt'; 
select_gnet_solver_def_file = 'solver/select_gnet_solver.prototxt';
caffe('init_gsolver', select_gnet_solver_def_file);
caffe('init_ssolver', select_snet_solver_def_file);
caffe('set_mode_gpu');

and thats the part that is called from matcaffe.cpp

static void init_gsolver(MEX_ARGS) {
  net_.reset();
  gsolver_.reset();
  //Step 1: init solver
  if (nrhs < 1 || nrhs > 3) {
    LOG(ERROR) << "Only given " << nrhs << " arguments";
    mexErrMsgTxt("caffe_mex : Wrong number of arguments");
  }

  char* solver_file = mxArrayToString(prhs[0]);

  LOG(INFO) << "Loading from " << solver_file;
  SolverParameter solver_param;
  ReadProtoFromTextFile(solver_file, &solver_param);

  gsolver_.reset(new SGDSolver<float>(solver_param));

  char* model_file;
  if (nrhs >1) {
    model_file = mxArrayToString(prhs[1]);
    if ( !string(model_file).empty() ){
      LOG(INFO) << "Recovery from " << model_file;
      gsolver_->net()->CopyTrainedLayersFrom(string(model_file));	
    }
    mxFree(model_file);
  }
  gsolver_->SetIter(0);
  mxFree(solver_file);
}

But I can't anything here...