fix generate_proposal_labels in cascade-rcnn series model, test=develop #27892
Conversation
|
Thanks for your contribution! |
|
✅ This PR's description meets the template requirements! |
jerrywgz
force-pushed
the
fix_generate_proposals_labels
branch
from
ba40645 to
8ed941c
Compare
… fix_generate_proposals_labels
| template <typename T> | ||
| void MaxOverlaps(const framework::Tensor& iou, | ||
| framework::Tensor* max_overlaps) { | ||
| const T* proposal_to_gt_overlaps = iou.data<T>(); |
There was a problem hiding this comment.
作为util的通用函数，命名也需要更通用。 proposal_to_gt_overlaps -> iou_data
| @@ -149,5 +149,19 @@ void ClipTiledBoxes(const platform::DeviceContext& ctx, | |||
| } | |||
| } | |||
|
|
|||
| template <typename T> | |||
| void MaxOverlaps(const framework::Tensor& iou, | |||
| framework::Tensor* max_overlaps) { | |||
There was a problem hiding this comment.
max_overlaps -> max_iou?
| const T* proposal_to_gt_overlaps = iou.data<T>(); | ||
| int row = iou.dims()[0]; | ||
| int col = iou.dims()[1]; | ||
| T* max_overlaps_dt = max_overlaps->data<T>(); |
There was a problem hiding this comment.
max_overlaps_dt -> max_iou_data
| @@ -2651,25 +2653,29 @@ def generate_proposal_labels(rpn_rois, | |||
| use_random(bool): Use random sampling to choose foreground and background boxes. | |||
| is_cls_agnostic(bool): bbox regression use class agnostic simply which only represent fg and bg boxes. | |||
| is_cascade_rcnn(bool): it will filter some bbox crossing the image's boundary when setting True. | |||
| max_overlap(Variable): Maximum overlap between each Input box and ground-truth. | |||
There was a problem hiding this comment.
Input box -> each proposal box .. ?
| @@ -2651,25 +2653,29 @@ def generate_proposal_labels(rpn_rois, | |||
| use_random(bool): Use random sampling to choose foreground and background boxes. | |||
| is_cls_agnostic(bool): bbox regression use class agnostic simply which only represent fg and bg boxes. | |||
| is_cascade_rcnn(bool): it will filter some bbox crossing the image's boundary when setting True. | |||
| max_overlap(Variable): Maximum overlap between each Input box and ground-truth. | |||
| return_max_overlap(bool): Whether return the maximum overlap between each Output box and ground-truth. | |||
| is_cascade_rcnn=False): | ||
| is_cascade_rcnn=False, | ||
| max_overlap=None, | ||
| return_max_overlap=False): |
There was a problem hiding this comment.
max_overlap是否始终return，而不加这个参数
There was a problem hiding this comment.
这里希望可以兼容PaddleDetection先前的版本，如果始终return，https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.4/ppdet/modeling/architectures/mask_rcnn.py#L125 计算 loss的位置会受到影响
| gt_boxes = fluid.data( | ||
| name='gt_boxes', shape=[6, 4], dtype='float32', lod_level=1) | ||
| im_info = fluid.data( | ||
| name='im_info', shape=[1, 3], dtype='float32', lod_level=1) |
| max_overlap_slice = | ||
| max_overlap->Slice(rpn_rois_lod[i], rpn_rois_lod[i + 1]); | ||
| } else { | ||
| max_overlap_slice.mutable_data<T>(im_info_slice.dims(), |
There was a problem hiding this comment.
这里的shape是和im_info_slice相同吗？ im_info_slice的shape是[1, 3]?
| AddInput("MaxOverlap", | ||
| "(LoDTensor), This input is a 1D LoDTensor with shape [N]." | ||
| "N is the number of Input(RpnRois), " | ||
| "each element is the maxoverlap between " |
There was a problem hiding this comment.
maxoverlap -> max overlap
| for (int i = 0; i < rois_num; ++i) { | ||
| if ((rpn_rois_dt[i * 4 + 2] - rpn_rois_dt[i * 4 + 0] + 1) > 0 && | ||
| (rpn_rois_dt[i * 4 + 3] - rpn_rois_dt[i * 4 + 1] + 1) > 0 && | ||
| max_overlap_dt[i] < 1.) { |
There was a problem hiding this comment.
add comments for why filter max_overlap_dt < 1.0
PR types
Function optimization
PR changes
OPs
Describe
Fix generate_proposal_labels in cascade-rcnn series model and add max_overlap as input.
This op is only used at the stage of training so it will have no effect on inference.
This op is only used in static mode so core.ops,xxx is not used as well.
The mAP of related models:
Cascade Faster RCNN is 40.8 and previous version is 40.9.
HTC is 42.9/37.0 and previous is 42.7/36.8
The difference on document is marked as below:
