Code for paper "Dynamic R-CNN: Towards High Quality Object Detection via Dynamic Training" #3040
Conversation
|
|
|
Hi @hkzhang95 |
| torch.topk( | ||
| assign_result.max_overlaps, | ||
| min(self.iou_topk, | ||
| len(assign_result.max_overlaps)))[0][-1].item()) |
There was a problem hiding this comment.
Simplify the logic to make it more readable for users.
iou_topk = min(self.iou_topk, len(assign_result.max_overlaps))
ious, _ = torch.topk(assign_result.max_overlaps, iou_topk)
cur_iou.append(ious[-1].item())
What do you mean by saying that? |
| self.beta_topk = self.train_cfg.dynamic_rcnn.beta_topk | ||
| self.iteration_count = self.train_cfg.dynamic_rcnn.iteration_count | ||
| self.initial_iou = 0.4 | ||
| self.initial_beta = 1.0 |
There was a problem hiding this comment.
Better to make initial_iou and initial_beta configurable.
There was a problem hiding this comment.
Added related settings in the configuration file.
| super(DynamicRoIHead, self).__init__(**kwargs) | ||
| self.iou_topk = self.train_cfg.dynamic_rcnn.iou_topk | ||
| self.beta_topk = self.train_cfg.dynamic_rcnn.beta_topk | ||
| self.iteration_count = self.train_cfg.dynamic_rcnn.iteration_count |
There was a problem hiding this comment.
iteration_count -> udpate_iter_interval
|
|
||
| def update_statistics(self): | ||
| new_iou_thr = max(self.initial_iou, | ||
| sum(self.cur_iou) / self.iteration_count) |
There was a problem hiding this comment.
new_iou_thr = max(self.initial_iou, np.mean(self. iou_history))
| self.initial_iou = 0.4 | ||
| self.initial_beta = 1.0 | ||
| self.cur_iou = [] | ||
| self.cur_beta = [] |
There was a problem hiding this comment.
Use more comprehensible variable names.
# the IoU history of the past `udpate_iter_interval` iterations
self.iou_history = []
# the beta history of the past `udpate_iter_interval` iterations
self.beta_history = []| bbox_results.update(loss_bbox=loss_bbox) | ||
| return bbox_results | ||
|
|
||
| def update_statistics(self): |
There was a problem hiding this comment.
update_statisticsis not a good method name since this method is not updating statistics but updating parameters like iou_thr and beta.- Add a docstring.
There was a problem hiding this comment.
Change the method name to update_hyperparameters and add a docstring.
| new_beta = min(self.initial_beta, | ||
| sorted(self.cur_beta)[self.iteration_count // 2]) | ||
| self.cur_beta = [] | ||
| assert isinstance(self.bbox_head.loss_bbox, SmoothL1Loss) |
There was a problem hiding this comment.
This assertion can be moved to __init__ method.
| self.bbox_assigner.neg_iou_thr = new_iou_thr | ||
| self.bbox_assigner.min_pos_iou = new_iou_thr | ||
| new_beta = min(self.initial_beta, | ||
| sorted(self.cur_beta)[self.iteration_count // 2]) |
There was a problem hiding this comment.
new_beta = min(self.initial_beta, np.median(self.beta_history))
…via Dynamic Training" (open-mmlab#3040) * Code for paper "Dynamic R-CNN: Towards High Quality Object Detection via Dynamic Training" * update configs/dynamic_rcnn/dynamic_rcnn_r50_fpn_1x.py * reformat code * simplify code * update model link * simplify code * simplify code logic * simplify code and add comments * minor updates of the docstring Co-authored-by: Jiarui XU <xvjiarui0826@gmail.com> Co-authored-by: Kai Chen <chenkaidev@gmail.com>
No description provided.