Fine-tune B-CNN models on cub failed

[Yili Zhao]

When invoked 'run run_experiments_bcnn_train.m' from MATLAB:

error use vl_argparse (line 99)
Unknown parameter 'cropSize'

error vl_argparse (line 79)
      opts = vl_argparse(opts, vertcat(params,values)) ;

error imdb_get_batch_bcnn (line 35)
    opts(i) = vl_argparse(opts(i), {varargin{1}(i),varargin{2:end}});

error getBatchSimpleNNWrapper>getBatchSimpleNN (line 10)
im = imdb_get_batch_bcnn(images, opts, 'prefetch', nargout == 0);

error getBatchSimpleNNWrapper>@(imdb,batch)getBatchSimpleNN(imdb,batch,opts) (line 4)
fn = @(imdb, batch) getBatchSimpleNN(imdb, batch, opts) ;

error initializeNetworkSharedWeights (line 117)
            [im, labels] = getBatchFn(imdb, batch) ;

error imdb_bcnn_train_dag (line 65)
net = initNetFn(imdb, encoderOpts, opts);

error run_experiments_bcnn_train (line 79)
            imdb_bcnn_train_dag(imdb, opts);

error run (line 96)
evalin('caller', [script ';']);

[tsungyu]

Did you modify the code? I didn't see any option "cropSize" in my implementation. You need to give a default value to the option If you want to add more. To be clear, before you pass opts to vl_argparse, 'cropSize' should be a field in opts.

[Yili Zhao]

For validating this error, I just do a clean clone and re-run 'run run_experiments_bcnn_train.m' again, but still got this error.

[Yili Zhao]

I setup a breakpoint at line 35 https://bitbucket.org/tsungyu/bcnn/src/e0a84034814a073738bdd208b876a02953f31805/imdb_get_batch_bcnn.m?at=master&fileviewer=file-view-default#imdb_get_batch_bcnn.m-35 and output the value of varargin{1}:

varargin{1}

ans = 

  struct:

         imageSize: [224 224 3 10]
        keepAspect: 1
      averageImage: [224×224×3 double]
            border: [32 32]
          cropSize: [0.8750 0.8750]
     interpolation: 'bilinear'
        numThreads: 12
    transformation: 'none'
       rgbVariance: []
             scale: 2

You can see there is a field named cropSize indeed, however, opts has no corresponding field. So maybe this why the vl_argparse error.

[Yili Zhao]

By the way, if add opts(i).cropSize = [1.0, 1.0]; at the line 35 https://bitbucket.org/tsungyu/bcnn/src/e0a84034814a073738bdd208b876a02953f31805/imdb_get_batch_bcnn.m?at=master&fileviewer=file-view-default#imdb_get_batch_bcnn.m-35, now I can run run_experiments_bcnn_train.m. However, there will be a segmentation fault like:

I test this on a small dataset. And if you want to validate it, I can send the dataset to you.

[tsungyu]

Seems the format of the matconvnet pre-trained models hs changed. I'll suggest to download the models in the previous version from here: http://www.vlfeat.org/matconvnet/models . The one in beta18 folder should work.

[Yili Zhao]

I think you are right. I check the net.meta structure, and find the cropSize field comes from the pre-trained model, which I download from matconvnet website, but it seems that this model has different format with the one trained from lower version of matconvnet. I will check it again.

[Yili Zhao]

The cropSize field problem was resolved by using the older model, however, the MATLAB System Error still exist. I don't know if it is caused by my software stack: matlab r2016b, cuda 8.0, cudnn 5.1 and matconvnet beta19. @tsungyu May you try the current repo with cub again to see if this segment fault exist?

[tsungyu]

I have tested my code using MatcConvnet beta19, Matlab 2014b, cuda 7.5 and cudnn5.0. I guess the cuda version might be the problem. Could you try downgrade the cuda version ans see if the error still exists?

[Yili Zhao]

My current GPU GTX 1070 depends on cuda 8.0, so I need to will find other hardware for testing. By the way, does bcnn support cpu training? I try to set useGpu to 0, but it seems that it doesn't work though.