Bump version to 2.0.0 (#2831)

docs/en/understand_mmcv/ops.md

@@ -9,7 +9,7 @@ We implement common ops used in detection, segmentation, etc.
 | BallQuery                    |     | √    | √   |     |        |
 | BBoxOverlaps                 |     | √    | √   | √   | √      |
 | BorderAlign                  |     | √    |     |     |        |
-| BoxIouRotated                | √   | √    | √   |     |        |
+| BoxIouRotated                | √   | √    | √   |     | √      |
 | BoxIouQuadri                 | √   | √    |     |     |        |
 | CARAFE                       |     | √    | √   |     |        |
 | ChamferDistance              |     | √    |     |     |        |

docs/zh_cn/understand_mmcv/ops.md

@@ -9,7 +9,7 @@ MMCV 提供了检测、分割等任务中常用的算子
 | BallQuery                    |     | √    | √   |     |        |
 | BBoxOverlaps                 |     | √    | √   | √   | √      |
 | BorderAlign                  |     | √    |     |     |        |
-| BoxIouRotated                | √   | √    | √   |     |        |
+| BoxIouRotated                | √   | √    | √   |     | √      |
 | BoxIouQuadri                 | √   | √    |     |     |        |
 | CARAFE                       |     | √    | √   |     |        |
 | ChamferDistance              |     | √    |     |     |        |

mmcv/ops/box_iou_rotated.py

@@ -142,6 +142,11 @@ def box_iou_rotated(bboxes1: torch.Tensor,
         flip_mat[-1] = -1
         bboxes1 = bboxes1 * flip_mat
         bboxes2 = bboxes2 * flip_mat
+    if bboxes1.device.type == 'npu':
+        scale_mat = bboxes1.new_ones(bboxes1.shape[-1])
+        scale_mat[-1] = 1.0 / 0.01745329252
+        bboxes1 = bboxes1 * scale_mat
+        bboxes2 = bboxes2 * scale_mat
     bboxes1 = bboxes1.contiguous()
     bboxes2 = bboxes2.contiguous()
     ext_module.box_iou_rotated(

mmcv/ops/csrc/pytorch/npu/box_iou_rotated_npu.cpp

@@ -0,0 +1,47 @@
+#include "pytorch_npu_helper.hpp"
+
+using namespace NPU_NAME_SPACE;
+using namespace std;
+
+void box_iou_rotated_impl(const Tensor boxes1, const Tensor boxes2, Tensor ious,
+                          const int mode_flag, const bool aligned);
+
+void box_iou_rotated_npu(const Tensor boxes1, const Tensor boxes2, Tensor ious,
+                         const int mode_flag, const bool aligned) {
+  at::Tensor boxes = at::ones_like(boxes1);
+  at::Tensor query_boxes = at::ones_like(boxes2);
+  boxes = boxes1.transpose(0, 1).unsqueeze(0);
+  query_boxes = boxes2.transpose(0, 1).unsqueeze(0);
+
+  bool is_trans = false;
+  string modeStr = "iou";
+  if (mode_flag == 1) {
+    modeStr = "iof";
+  }
+  bool is_cross = true;
+  if (aligned) {
+    is_cross = false;
+  }
+  float v_threshold = 0;
+  float e_threshold = 0;
+
+  OpCommand cmd;
+  cmd.Name("RotatedIou")
+      .Input(boxes)
+      .Input(query_boxes)
+      .Output(ious)
+      .Attr("trans", is_trans)
+      .Attr("mode", modeStr)
+      .Attr("is_cross", is_cross)
+      .Attr("v_threshold", v_threshold)
+      .Attr("e_threshold", e_threshold)
+      .Run();
+
+  if (is_cross) {
+    ious = ious.view({boxes1.size(0), boxes2.size(0)});
+  } else {
+    ious = ious.view({boxes1.size(0), 1});
+  }
+}
+
+REGISTER_NPU_IMPL(box_iou_rotated_impl, box_iou_rotated_npu);

tests/test_ops/test_box_iou_rotated.py

@@ -4,7 +4,7 @@
 import torch
 
 from mmcv.ops import box_iou_rotated
-from mmcv.utils import IS_CUDA_AVAILABLE, IS_MLU_AVAILABLE
+from mmcv.utils import IS_CUDA_AVAILABLE, IS_MLU_AVAILABLE, IS_NPU_AVAILABLE
 
 
 class TestBoxIoURotated:
@@ -54,7 +54,11 @@ def test_box_iou_rotated_cpu(self):
         pytest.param(
             'mlu',
             marks=pytest.mark.skipif(
-                not IS_MLU_AVAILABLE, reason='requires MLU support'))
+                not IS_MLU_AVAILABLE, reason='requires MLU support')),
+        pytest.param(
+            'npu',
+            marks=pytest.mark.skipif(
+                not IS_NPU_AVAILABLE, reason='requires NPU support'))
     ])
     def test_box_iou_rotated(self, device):
         np_boxes1 = np.asarray(
@@ -76,14 +80,27 @@ def test_box_iou_rotated(self, device):
         boxes2 = torch.from_numpy(np_boxes2).to(device)
 
         # test cw angle definition
+        # gpu
         ious = box_iou_rotated(boxes1, boxes2)
         assert np.allclose(ious.cpu().numpy(), np_expect_ious, atol=1e-4)
 
         ious = box_iou_rotated(boxes1, boxes2, aligned=True)
         assert np.allclose(
             ious.cpu().numpy(), np_expect_ious_aligned, atol=1e-4)
 
+        # npu
+        boxes1_npu = torch.from_numpy(np_boxes1).npu()
+        boxes2_npu = torch.from_numpy(np_boxes2).npu()
+
+        ious = box_iou_rotated(boxes1_npu, boxes2_npu)
+        assert np.allclose(ious.cpu().numpy(), np_expect_ious, atol=1e-4)
+
+        ious = box_iou_rotated(boxes1_npu, boxes2_npu, aligned=True)
+        assert np.allclose(
+            ious.cpu().numpy(), np_expect_ious_aligned, atol=1e-4)
+
         # test ccw angle definition
+        # gpu
         boxes1[..., -1] *= -1
         boxes2[..., -1] *= -1
         ious = box_iou_rotated(boxes1, boxes2, clockwise=False)
@@ -93,6 +110,17 @@ def test_box_iou_rotated(self, device):
         assert np.allclose(
             ious.cpu().numpy(), np_expect_ious_aligned, atol=1e-4)
 
+        # npu
+        boxes1_npu[..., -1] *= -1
+        boxes2_npu[..., -1] *= -1
+        ious = box_iou_rotated(boxes1_npu, boxes2_npu, clockwise=False)
+        assert np.allclose(ious.cpu().numpy(), np_expect_ious, atol=1e-4)
+
+        ious = box_iou_rotated(
+            boxes1_npu, boxes2_npu, aligned=True, clockwise=False)
+        assert np.allclose(
+            ious.cpu().numpy(), np_expect_ious_aligned, atol=1e-4)
+
     def test_box_iou_rotated_iof_cpu(self):
         np_boxes1 = np.asarray(
             [[1.0, 1.0, 3.0, 4.0, 0.5], [2.0, 2.0, 3.0, 4.0, 0.6],
@@ -137,7 +165,11 @@ def test_box_iou_rotated_iof_cpu(self):
         pytest.param(
             'mlu',
             marks=pytest.mark.skipif(
-                not IS_MLU_AVAILABLE, reason='requires MLU support'))
+                not IS_MLU_AVAILABLE, reason='requires MLU support')),
+        pytest.param(
+            'npu',
+            marks=pytest.mark.skipif(
+                not IS_NPU_AVAILABLE, reason='requires NPU support'))
     ])
     def test_box_iou_rotated_iof(self, device):
         np_boxes1 = np.asarray(
@@ -159,14 +191,28 @@ def test_box_iou_rotated_iof(self, device):
         boxes2 = torch.from_numpy(np_boxes2).to(device)
 
         # test cw angle definition
+        # gpu
         ious = box_iou_rotated(boxes1, boxes2, mode='iof')
         assert np.allclose(ious.cpu().numpy(), np_expect_ious, atol=1e-4)
 
         ious = box_iou_rotated(boxes1, boxes2, mode='iof', aligned=True)
         assert np.allclose(
             ious.cpu().numpy(), np_expect_ious_aligned, atol=1e-4)
 
+        # npu
+        boxes1_npu = torch.from_numpy(np_boxes1).npu()
+        boxes2_npu = torch.from_numpy(np_boxes2).npu()
+
+        ious = box_iou_rotated(boxes1_npu, boxes2_npu, mode='iof')
+        assert np.allclose(ious.cpu().numpy(), np_expect_ious, atol=1e-4)
+
+        ious = box_iou_rotated(
+            boxes1_npu, boxes2_npu, mode='iof', aligned=True)
+        assert np.allclose(
+            ious.cpu().numpy(), np_expect_ious_aligned, atol=1e-4)
+
         # test ccw angle definition
+        # gpu
         boxes1[..., -1] *= -1
         boxes2[..., -1] *= -1
         ious = box_iou_rotated(boxes1, boxes2, mode='iof', clockwise=False)
@@ -176,3 +222,15 @@ def test_box_iou_rotated_iof(self, device):
             boxes1, boxes2, mode='iof', aligned=True, clockwise=False)
         assert np.allclose(
             ious.cpu().numpy(), np_expect_ious_aligned, atol=1e-4)
+
+        # npu
+        boxes1_npu[..., -1] *= -1
+        boxes2_npu[..., -1] *= -1
+        ious = box_iou_rotated(
+            boxes1_npu, boxes2_npu, mode='iof', clockwise=False)
+        assert np.allclose(ious.cpu().numpy(), np_expect_ious, atol=1e-4)
+
+        ious = box_iou_rotated(
+            boxes1_npu, boxes2_npu, mode='iof', aligned=True, clockwise=False)
+        assert np.allclose(
+            ious.cpu().numpy(), np_expect_ious_aligned, atol=1e-4)