Bounding box model returns negative values for bbox

Hi, I re-implemented the bounding box models and fcn models from lesson 7 in keras + TF and got fairly good results on the fisheries example (though somewhat worse than Jeremy’s benchmark). In spite of this, I noticed that the bounding box values the model returned contained a substantial amount of negative values, while the minimal value in the bbox data it was trained with was zero (as it has to be).
I might be missing something with setting lower and upper bounds for this in keras, but couldn’t find anything. Or do you need to do any data normalization on the training bbox data? Didn’t find anything like this in lesson7.ipynb

Did anyone experience the same?