search

aleksispi / drl-rpn-tf

Official Tensorflow implementation of drl-RPN: Deep Reinforcement Learning of Region Proposal Networks (CVPR 2018 paper)
MIT License
77 stars 21 forks source link

Joint Training of Policy and Detector problem #5

Closed chang010453 closed 4 years ago

chang010453 commented 4 years ago

In your paper's section 5.3, it says theta_base and theta_det would be trained. In your code, they training at train_det mode. But I look into the gradients it produce, only theta_det can be update, there is no gradient in theta_base. Can you check this problem? thanks.

aleksispi commented 4 years ago

Hi,

The code does not support training of theta_base. See page 5 in the paper http://openaccess.thecvf.com/content_cvpr_2018/papers/Pirinen_Deep_Reinforcement_Learning_CVPR_2018_paper.pdf, where in the footnote 5 we write that theta_base is frozen as we saw no improvement by training those parameters.

If you would like to investigate also training the full system, probably a good place to start is the Faster R-CNN code that I built drl-RPN on top of. That code base should support training theta_base too: https://github.com/endernewton/tf-faster-rcnn. Hopefully you should be able to mimic the code they use for training theta_base and insert to the drl-RPN code.

Cheers, Aleksis