Improve NMS speed (#3467)

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>

ultralytics/yolo/utils/ops.py

@@ -200,12 +200,16 @@ def non_max_suppression(
     multi_label &= nc > 1  # multiple labels per box (adds 0.5ms/img)
     merge = False  # use merge-NMS
 
+    prediction = prediction.clone()  # don't modify original
+    prediction = prediction.transpose(-1, -2)  # to (batch, boxes, items)
+    prediction[..., :4] = xywh2xyxy(prediction[..., :4])  # xywh to xyxy
+
     t = time.time()
     output = [torch.zeros((0, 6 + nm), device=prediction.device)] * bs
     for xi, x in enumerate(prediction):  # image index, image inference
         # Apply constraints
         # x[((x[:, 2:4] < min_wh) | (x[:, 2:4] > max_wh)).any(1), 4] = 0  # width-height
-        x = x.transpose(0, -1)[xc[xi]]  # confidence
+        x = x[xc[xi]]  # confidence
 
         # Cat apriori labels if autolabelling
         if labels and len(labels[xi]):
@@ -221,9 +225,9 @@ def non_max_suppression(
 
         # Detections matrix nx6 (xyxy, conf, cls)
         box, cls, mask = x.split((4, nc, nm), 1)
-        box = xywh2xyxy(box)  # center_x, center_y, width, height) to (x1, y1, x2, y2)
+
         if multi_label:
-            i, j = (cls > conf_thres).nonzero(as_tuple=False).T
+            i, j = torch.where(cls > conf_thres)
             x = torch.cat((box[i], x[i, 4 + j, None], j[:, None].float(), mask[i]), 1)
         else:  # best class only
             conf, j = cls.max(1, keepdim=True)
@@ -241,6 +245,8 @@ def non_max_suppression(
         n = x.shape[0]  # number of boxes
         if not n:  # no boxes
             continue
+
+        if n > max_nms:  # excess boxes
             x = x[x[:, 4].argsort(descending=True)[:max_nms]]  # sort by confidence and remove excess boxes
 
         # Batched NMS