Open ishacusp opened 5 years ago
Thanks! I followed your README and used your retinanet.py and feature extractor, to train on my data, but the feature extractor returns an empty map.
@ishacusp Okay... I get it. This error is caused by inexplicit parameter assignment during retinanet feature extractor initialization when I add min_level
and max_level
in base class only. I have fixed this bug and you could pull it down to use. Thanks for reporting.
Thank You!
@ishacusp TF object detection has its own focal loss function integrated into its framework and you could enable it by defining
weighted_sigmoid_focal
in config loss section. If you'd like to use this customed loss function, modifyobject_detection/meta_architecture/ssd_meta_arch
to replace the default classification loss (Line 844).