Detectron.pytorch icon indicating copy to clipboard operation
Detectron.pytorch copied to clipboard

Published 20 hours ago • verified publisher icon roytseng-tw

support RPN proposals testing

Open w-hc opened this issue 7 years ago • 12 comments
trafficstars

Support single GPU and multi GPU testing on RPN proposals

Do not support RPN only training. Cases such as

  1. training only the RPN branch
  2. training on pre-computed proposals

require restructuring model_builder.py quite a bit. @roytseng-tw you should call the shots

w-hc avatar Aug 11 '18 00:08 w-hc

Hi, thanks for your work.

But I did not see these codes in the master branch?

qinhaifangpku avatar Oct 06 '18 02:10 qinhaifangpku

This pull request is not merged.

w-hc avatar Oct 06 '18 02:10 w-hc

This pull request is not merged. I see. How I can use them? Cause I want to generate the proposal from a mask rcnn trained model.

qinhaifangpku avatar Oct 06 '18 02:10 qinhaifangpku

This pull request is not merged.

and then train the model from the precompute proposals.

qinhaifangpku avatar Oct 06 '18 02:10 qinhaifangpku

https://stackoverflow.com/questions/6022302/how-to-apply-unmerged-upstream-pull-requests-from-other-forks-into-my-fork

Maybe this can help.

w-hc avatar Oct 06 '18 02:10 w-hc

https://stackoverflow.com/questions/6022302/how-to-apply-unmerged-upstream-pull-requests-from-other-forks-into-my-fork

Maybe this can help.

Okay, Thanks for your kindly help. I saw your repository has many branches. I wonder which branch and commit should I merge to the roys' master branch will make the rpn-testing work?

qinhaifangpku avatar Oct 06 '18 02:10 qinhaifangpku

I created a separate branch just for this pull request. It's called rpn-testing screen shot 2018-10-05 at 21 41 08

w-hc avatar Oct 06 '18 02:10 w-hc

I created a separate branch just for this pull request. It's called testing screen shot 2018-10-05 at 21 41 08

I see.

Thanks again!

qinhaifangpku avatar Oct 06 '18 02:10 qinhaifangpku

I created a separate branch just for this pull request. It's called rpn-testing screen shot 2018-10-05 at 21 41 08

Hi, I wonder if the model can be trained with the precomputed proposals and just train the mask branch and not update the box branch?

qinhaifangpku avatar Oct 08 '18 09:10 qinhaifangpku

This codebase does not support Fast-RCNN training (the one you described) or RPN-only training. Both require restructuring model_builder.py. I did not do this pull request since it is better for the author himself to decide on this.

Also, _compute_targets in roidb.py has a behavior difference compared to fast rcnn assignment policy regarding crowd. I would say it involves a bit of testing to make sure there is no bug.

w-hc avatar Oct 08 '18 14:10 w-hc

Support single GPU and multi GPU testing on RPN proposals

Do not support RPN only training. Cases such as

  1. training only the RPN branch
  2. training on pre-computed proposals

require restructuring model_builder.py quite a bit. @roytseng-tw you should call the shots

Hi, I wonder if there we could test the bbox and mask by given proposals?

Thanks for your help in advance!

qinhaifangpku avatar Oct 09 '18 08:10 qinhaifangpku

Some slight modifications will make it possible. screen shot 2018-10-09 at 11 15 34 You might need to sprinkle a few more changes here and there down the road. But the spirit is true.

w-hc avatar Oct 09 '18 16:10 w-hc