kentaroy47 / frcnn-from-scratch-with-keras

:collision:Faster R-CNN from scratch written with Keras
Apache License 2.0
168 stars 106 forks source link

RPN only but what about the classifier #26

Open b-k-nielsen opened 5 years ago

b-k-nielsen commented 5 years ago

Hi, thanks for the work. I'm using a similar implementation and think training the RPN separately is fine. However I don't understand that we train the complete pre-trained network. Wouldn't we like to keep that stable and train the classifier on top only? Or only train/replace the uppermost layers of the pre-trained network holding the complex features? In my trainings I've realized that the RPN degrades rather quickly like after about 100 epochs. Is this normal/explainable? Cheers, BKN