Dataset wrapper

[botcs]

Wrapper

In cases when multiple datasets are combined to train the network, we may end up doing a lot of extra work finding out which index refers to which class (e.g. training on CityScapes with keeping the ROI heads from a COCO pre-training as in #259).

This auxiliary dataset helps to find common class names between
    `mimicked_dataset` and `wrapped_dataset` and wrap the latter to return
    ground truth that aligns with the indices of the former.

        A: mimicked_dataset
        B: wrapped_dataset

When multiple dataset are provided in a config file, the dataset builder checks if the datasets belong to different classes, and if possible it will match the 2nd, 3rd ... nth dataset's categories to the 1st dataset's in such a way that the corresponding class indices will be identical to the 1st dataset.
Instances belonging to classes that were left unmatched by this process are discarded on the fly.

The dataset comes with a pretty printing __str__ that helps debugging.

Example output

[WrapperDataset mimicks:Sim10kDataset wraps:CityScapesDataset]
 Mimicked index : Mimicked label       Wrapped index : Wrapped label  
              0 : __background__  ->               0 : __background__ 
              1 : aeroplane       ->            None : __unmatched__(aeroplane)
              2 : bicycle         ->              10 : bicycle        
              3 : bird            ->            None : __unmatched__(bird)
              4 : boat            ->            None : __unmatched__(boat)
              5 : bottle          ->            None : __unmatched__(bottle)
              6 : bus             ->               5 : bus            
              7 : car             ->               3 : car            
              8 : cat             ->            None : __unmatched__(cat)
              9 : chair           ->            None : __unmatched__(chair)
             10 : cow             ->            None : __unmatched__(cow)
             11 : diningtable     ->            None : __unmatched__(diningtable)
             12 : dog             ->            None : __unmatched__(dog)
             13 : horse           ->            None : __unmatched__(horse)
             14 : motorbike       ->            None : __unmatched__(motorbike)
             15 : person          ->               1 : person         
             16 : pottedplant     ->            None : __unmatched__(pottedplant)
             17 : sheep           ->            None : __unmatched__(sheep)
             18 : sofa            ->            None : __unmatched__(sofa)
             19 : train           ->               8 : train          
             20 : tvmonitor       ->            None : __unmatched__(tvmonitor)