improve stability of test_nms_cuda (#2044)

* improve stability of test_nms_cuda

This change addresses two issues:

_create_tensors_with_iou() creates test data for the NMS tests. It
takes care to ensure at least one pair of boxes (1st and last) have
IoU around the threshold for the test. However, the constructed
IoU for that pair is _so_ close to the threshold that rounding
differences (presumably) between CPU and CUDA implementations may
result in one suppressing a box in the pair and the other not.
Adjust the construction to ensure the IoU for the box pair is
near the threshold, but far-enough above that both implementations
should agree.

Where 2 boxes have nearly or exactly the same score, the CPU and
CUDA implementations may order them differently. Adjust
test_nms_cuda() to check only that the non-suppressed box lists
include the same members, without regard for ordering.

* adjust assertion in test_nms_cuda

The CPU and CUDA nms implementations each sort the box scores
as part of their work, but the sorts they use are not stable. So
boxes with the same score maybe be processed in opposite order
by the two implmentations.

Relax the assertion in test_nms_cuda (following the model in
pytorch's test_topk()) to allow the test to pass if the output
differences are caused by similarly-scored boxes.

* improve stability of test_nms_cuda

Adjust _create_tensors_with_iou() to ensure we create at least
one box just over threshold that should be suppressed.

test/test_ops.py

@@ -374,10 +374,14 @@ class NMSTester(unittest.TestCase):
         # let b0 be [x0, y0, x1, y1], and b1 be [x0, y0, x1 + d, y1],
         # then, in order to satisfy ops.iou(b0, b1) == iou_thresh,
         # we need to have d = (x1 - x0) * (1 - iou_thresh) / iou_thresh
+        # Adjust the threshold upward a bit with the intent of creating
+        # at least one box that exceeds (barely) the threshold and so
+        # should be suppressed.
         boxes = torch.rand(N, 4) * 100
         boxes[:, 2:] += boxes[:, :2]
         boxes[-1, :] = boxes[0, :]
         x0, y0, x1, y1 = boxes[-1].tolist()
+        iou_thresh += 1e-5
         boxes[-1, 2] += (x1 - x0) * (1 - iou_thresh) / iou_thresh
         scores = torch.rand(N)
         return boxes, scores
@@ -399,7 +403,12 @@ class NMSTester(unittest.TestCase):
             r_cpu = ops.nms(boxes, scores, iou)
             r_cuda = ops.nms(boxes.cuda(), scores.cuda(), iou)
 
-            self.assertTrue(torch.allclose(r_cpu, r_cuda.cpu()), err_msg.format(iou))
+            is_eq = torch.allclose(r_cpu, r_cuda.cpu())
+            if not is_eq:
+                # if the indices are not the same, ensure that it's because the scores
+                # are duplicate
+                is_eq = torch.allclose(scores[r_cpu], scores[r_cuda.cpu()])
+            self.assertTrue(is_eq, err_msg.format(iou))
 
 
 class NewEmptyTensorTester(unittest.TestCase):