Training with class label subset

[ChrisAlthaus]

❓ Questions and Help

Hi, I wonder if it's possible to train the SGDet Model with custom class labels (subset of the original 150 labels) only to improve model performance just on this set of class labels? So far i can image to implement this by 2 methods:

1. Filter all box & relation annotations of the scene graphs in vg/VG-SGG-with-attri.h5
2. Modify load_graphs in /Scene-Graph-Benchmark.pytorch/maskrcnn_benchmark/data/datasets/visual_genome.py to filter out all unneeded class labels. (This approach i am implemented at the moment, but I am getting IndexError in get_groundtruth)

Thanks for this well documented repository and I'm looking forward for any suggestions.

[pcyyo]

have you solved this problem？