[fbsync] [OPS, IMP] New batched_nms implementation (#3426)

Summary:
* new batched_nms implem

* flake8

* hopefully fix torchscipt tests

* Use where instead of nonzero

* Use same threshold (4k) for CPU and GPU

* Remove use of argsort

* use views again

* remove print

* trying stuff, I don't know what's going on

* previous passed onnx checks so the error isn't in _vanilla func. Trying to return vanilla now

* add tracing decorators

* cleanup

* wip

* ignore new path with ONNX

* use vanilla if tracing...????

* Remove script_if_tracing decorator as it was conflicting with _is_tracing

* flake8

* Improve coverage

Reviewed By: NicolasHug, cpuhrsch

Differential Revision: D26945728

fbshipit-source-id: 118a41e03da2939a726e5bd18f5f77b7c0ce6339

Co-authored-by: Francisco Massa <[email protected]>

Author: fmassa

test/test_onnx.py

@@ -1,3 +1,11 @@
+# onnxruntime requires python 3.5 or above
+try:
+    # This import should be before that of torch
+    # see https://github.com/onnx/onnx/issues/2394#issuecomment-581638840
+    import onnxruntime
+except ImportError:
+    onnxruntime = None
+
 from common_utils import set_rng_seed
 import io
 import torch
@@ -13,12 +21,6 @@
 
 from collections import OrderedDict
 
-# onnxruntime requires python 3.5 or above
-try:
-    import onnxruntime
-except ImportError:
-    onnxruntime = None
-
 import unittest
 from torchvision.ops._register_onnx_ops import _onnx_opset_version
 

test/test_ops.py

@@ -461,6 +461,28 @@ def test_nms_cuda_float16(self):
         keep16 = ops.nms(boxes.to(torch.float16), scores.to(torch.float16), iou_thres)
         self.assertTrue(torch.all(torch.eq(keep32, keep16)))
 
+    def test_batched_nms_implementations(self):
+        """Make sure that both implementations of batched_nms yield identical results"""
+
+        num_boxes = 1000
+        iou_threshold = .9
+
+        boxes = torch.cat((torch.rand(num_boxes, 2), torch.rand(num_boxes, 2) + 10), dim=1)
+        assert max(boxes[:, 0]) < min(boxes[:, 2])  # x1 < x2
+        assert max(boxes[:, 1]) < min(boxes[:, 3])  # y1 < y2
+
+        scores = torch.rand(num_boxes)
+        idxs = torch.randint(0, 4, size=(num_boxes,))
+        keep_vanilla = ops.boxes._batched_nms_vanilla(boxes, scores, idxs, iou_threshold)
+        keep_trick = ops.boxes._batched_nms_coordinate_trick(boxes, scores, idxs, iou_threshold)
+
+        err_msg = "The vanilla and the trick implementation yield different nms outputs."
+        self.assertTrue(torch.allclose(keep_vanilla, keep_trick), err_msg)
+
+        # Also make sure an empty tensor is returned if boxes is empty
+        empty = torch.empty((0,), dtype=torch.int64)
+        self.assertTrue(torch.allclose(empty, ops.batched_nms(empty, None, None, None)))
+
 
 class DeformConvTester(OpTester, unittest.TestCase):
     def expected_fn(self, x, weight, offset, mask, bias, stride=1, padding=0, dilation=1):

torchvision/ops/boxes.py

@@ -1,6 +1,6 @@
 import torch
 from torch import Tensor
-from typing import Tuple
+from typing import List, Tuple
 from ._box_convert import _box_cxcywh_to_xyxy, _box_xyxy_to_cxcywh, _box_xywh_to_xyxy, _box_xyxy_to_xywh
 import torchvision
 from torchvision.extension import _assert_has_ops
@@ -36,7 +36,6 @@ def nms(boxes: Tensor, scores: Tensor, iou_threshold: float) -> Tensor:
     return torch.ops.torchvision.nms(boxes, scores, iou_threshold)
 
 
-@torch.jit._script_if_tracing
 def batched_nms(
     boxes: Tensor,
     scores: Tensor,
@@ -62,18 +61,50 @@ def batched_nms(
             the elements that have been kept by NMS, sorted
             in decreasing order of scores
     """
-    if boxes.numel() == 0:
-        return torch.empty((0,), dtype=torch.int64, device=boxes.device)
-    # strategy: in order to perform NMS independently per class.
+    # Benchmarks that drove the following thresholds are at
+    # https://github.com/pytorch/vision/issues/1311#issuecomment-781329339
+    # Ideally for GPU we'd use a higher threshold
+    if boxes.numel() > 4_000 and not torchvision._is_tracing():
+        return _batched_nms_vanilla(boxes, scores, idxs, iou_threshold)
+    else:
+        return _batched_nms_coordinate_trick(boxes, scores, idxs, iou_threshold)
+
+
+@torch.jit._script_if_tracing
+def _batched_nms_coordinate_trick(
+    boxes: Tensor,
+    scores: Tensor,
+    idxs: Tensor,
+    iou_threshold: float,
+) -> Tensor:
+    # strategy: in order to perform NMS independently per class,
     # we add an offset to all the boxes. The offset is dependent
     # only on the class idx, and is large enough so that boxes
     # from different classes do not overlap
-    else:
-        max_coordinate = boxes.max()
-        offsets = idxs.to(boxes) * (max_coordinate + torch.tensor(1).to(boxes))
-        boxes_for_nms = boxes + offsets[:, None]
-        keep = nms(boxes_for_nms, scores, iou_threshold)
-        return keep
+    if boxes.numel() == 0:
+        return torch.empty((0,), dtype=torch.int64, device=boxes.device)
+    max_coordinate = boxes.max()
+    offsets = idxs.to(boxes) * (max_coordinate + torch.tensor(1).to(boxes))
+    boxes_for_nms = boxes + offsets[:, None]
+    keep = nms(boxes_for_nms, scores, iou_threshold)
+    return keep
+
+
+@torch.jit._script_if_tracing
+def _batched_nms_vanilla(
+    boxes: Tensor,
+    scores: Tensor,
+    idxs: Tensor,
+    iou_threshold: float,
+) -> Tensor:
+    # Based on Detectron2 implementation, just manually call nms() on each class independently
+    keep_mask = torch.zeros_like(scores, dtype=torch.bool)
+    for class_id in torch.unique(idxs):
+        curr_indices = torch.where(idxs == class_id)[0]
+        curr_keep_indices = nms(boxes[curr_indices], scores[curr_indices], iou_threshold)
+        keep_mask[curr_indices[curr_keep_indices]] = True
+    keep_indices = torch.where(keep_mask)[0]
+    return keep_indices[scores[keep_indices].sort(descending=True)[1]]
 
 
 def remove_small_boxes(boxes: Tensor, min_size: float) -> Tensor: