Several errors encountered attempting to train on a single class VOC 2007

Hello. After a week or so of trial and error i have to resort to you smart people to help me out in my attempt to fine-tune F-RCNN on one VOC class.

This issue regards running "script_faster_rcnn_VOC2007_ZF.m" but applies to any dataset. To replicate the original work, i use MATLAB 2014a with CUDA 6.5 on a gtx1080. To start off, i manage to train on the original dataset, with all classes without issues.

Following 1 i reduced the following line in VOCinit.m:

VOCopts.classes={... 'boat'};

and in .\models\fast_rcnn_prototxts\ZF_fc6\train_val.prototxt:

input: "bbox_targets"
input_dim: 1  # to be changed on-the-fly to match num ROIs
input_dim: 8 # 4 * (K+1) (=1) class

input: "bbox_loss_weights"
input_dim: 1  # to be changed on-the-fly to match num ROIs
input_dim: 8 # 4 * (K+1) (=1) class

layer {
    bottom: "fc7"
    top: "cls_score"
    name: "cls_score" ...

    inner_product_param {
        num_output: 2

According to aforementioned issue 1, this should be sufficient to begin training. However, as the annotation-files still contain all classes, thus i run into this MATLAB error while building roidb cache:

roidb (voc_2007_trainval): 1/5011
Error using containers.Map/values
The specified key is not present in this container.

Error in roidb_from_voc>attach_proposals (line 182)
  gt_classes =
  class_to_id.values({voc_rec.objects(valid_objects).class});

I have also here tried to change .\datasets\VOCdevkit2007\VOC2007\ImageSets\Main\test.txt ..\trainval.txt ..\val.txt ..\train.txt to match what is contained in ..\boat_test.txt ..\boat_trainval.txt ..\boat_val.txt ..\boat_test.txt. As these images still contain all the other classes it is to no avail, and the error is still present.

To conquer this, i made a new set of annotations, where i removed all the

ShaoqingRen / faster_rcnn

Several errors encountered attempting to train on a single class VOC 2007 #150