Is there a bug in the OHEMSampler? #3677
Conversation
fix the bug in the ohem_sampler
|
|
Codecov Report
@@ Coverage Diff @@
## master #3677 +/- ##
==========================================
- Coverage 61.10% 61.06% -0.05%
==========================================
Files 216 216
Lines 15222 15318 +96
Branches 2530 2608 +78
==========================================
+ Hits 9302 9354 +52
- Misses 5458 5498 +40
- Partials 462 466 +4
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
| self.bbox_roi_extractor = context.bbox_roi_extractor[ | ||
| context.current_stage] | ||
| self.bbox_head = context.bbox_head[context.current_stage] | ||
| self.bbox_head = self.context.bbox_head[self.context.current_stage] |
There was a problem hiding this comment.
There is something wrong with the OHEM usage in cascade rcnn. May I ask if you would like to fix that? If not, you can add an assertion here and I will fix it soon.
There was a problem hiding this comment.
There is something wrong with the OHEM usage in cascade rcnn. May I ask if you would like to fix that? If not, you can add an assertion here and I will fix it soon.
Yes, I'd like to fix that, but I only have time on weekends.
There was a problem hiding this comment.
No problem
I fixed the OHEM usage problem in cascade rcnn，please have a check
I got the error(TypeError: forward() missing 1 required positional argument: 'x_reg') when i used the OHEMSampler in the DoublehandRCNN, so I checkd the ohem_sampler.py and double_roi_head.py.
line 27(double_roi_head.py) :
cls_score, bbox_pred = self.bbox_head(bbox_cls_feats, bbox_reg_feats)
line 37(ohem_sampler.py):
cls_score, _ = self.bbox_head(bbox_feats)
the arguments are different.
in the meantime, self.with_shared_head is in the double_roi_head.py but not in the ohem_sampler.py
so I did some fixes