Add nms op and batched_nms api #40962
Conversation
| Tensor* output = context.Output<Tensor>("KeepBoxesIdxs"); | ||
| int64_t* output_data = output->mutable_data<int64_t>(context.GetPlace()); | ||
| auto threshold = context.template Attr<float>("iou_threshold"); | ||
| VLOG(3) << "threshold= " << threshold; |
|
|
||
| static const int64_t threadsPerBlock = sizeof(int64_t) * 8; | ||
|
|
||
| __host__ __device__ static inline int64_t CeilDivide(int64_t n, int64_t m) { |
There was a problem hiding this comment.
"host device" -> HOSTDEVICE
… add_nms_op_and_batched_nms 1. Make current branch updated. 2. Try to solve CI "build Kunlun KP build error"
| "Boxes is a Tensor with shape [N, 4] " | ||
| "N is the number of boxes " | ||
| "in last dimension in format [x1, x2, y1, y2] " | ||
| "the relation shold be ``0 <= x1 < x2 && 0 <= y1 < y2``."); |
|
|
||
| HOSTDEVICE static inline int64_t CeilDivide(int64_t n, int64_t m) { | ||
| return (n + m - 1) / m; | ||
| } |
There was a problem hiding this comment.
This common function can be put in nms_op.h, then remove this function in .cc
| template <typename T> | ||
| static __device__ inline bool CalculateIoU(const T* const box_1, | ||
| const T* const box_2, | ||
| const float threshold) { |
| else: | ||
| continue | ||
|
|
||
| return selected_indices |
There was a problem hiding this comment.
Remove duplicated code, import from other file
python/paddle/vision/ops.py
Outdated
| return out | ||
|
|
||
|
|
||
| def batched_nms(boxes, scores, category_idxs, categories, iou_threshold, top_k): |
There was a problem hiding this comment.
please also add unit testing for Dynamic to Static
There was a problem hiding this comment.
added test_batched_nms_dynamic_to_static function in python/paddle/fluid/tests/unittest/test_batched_nms.py
python/paddle/vision/ops.py
Outdated
| return out | ||
|
|
||
|
|
||
| def batched_nms(boxes, scores, category_idxs, categories, iou_threshold, top_k): |
There was a problem hiding this comment.
nms和batched_nms是否是业界普遍的叫法？
batched_nms和nms这两个api是否能在对外接口层合并？比如，默认category_idxs=None，表示nms的情况
- 一般api不区分batched和非batched的形式，非batched认为是batched一种特殊形式，比如不会有batched_conv
- batched一般理解是输入数据第0维度表示样本的batch，但这里的含义并不相同
There was a problem hiding this comment.
当时使用该名称时仅参考了pytorch的设计，经调研batched_nms仅在pytorch中使用。与指导人讨论后认为batched一次确实无法较好概括该API所实现的功能。目前已经将这两个API合并为同一个nms，使用参数进行区分。
… add_nms_op_and_batched_nms merge to modify unittest CMakefile.txt to skip test_ops_nms
| std::vector<uint64_t> mask_host(num_boxes * blocks_per_line); | ||
| memory::Copy(platform::CPUPlace(), mask_host.data(), context.GetPlace(), | ||
| mask_dev, num_boxes * blocks_per_line * sizeof(uint64_t), | ||
| context.cuda_device_context().stream()); |
There was a problem hiding this comment.
GPU内容拷回CPU后，需要同步，不然后面用到的mask_host极有可能是脏数据。
PR types
New features
PR changes
OPs
Describe
add nms op in paddle/fluid/operators/detection and corresponding api batched_nms In python/paddle/vision/ops.py
English document preview:
中文文档预览：
