You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: configs/ppyolo/README.md
+2-2Lines changed: 2 additions & 2 deletions
Original file line number
Diff line number
Diff line change
@@ -70,7 +70,7 @@ PP-YOLO improved performance and speed of YOLOv3 with following methods:
70
70
71
71
**Notes:**
72
72
73
-
- PP-YOLO_MobileNetV3 is trained on COCO train2017 datast and evaluated on val2017 dataset,Box AP<sup>val</sup> is evaluation results of `mAP(IoU=0.5:0.95)`, Box AP<sup>val</sup> is evaluation results of `mAP(IoU=0.5)`.
73
+
- PP-YOLO_MobileNetV3 is trained on COCO train2017 datast and evaluated on val2017 dataset,Box AP<sup>val</sup> is evaluation results of `mAP(IoU=0.5:0.95)`, Box AP50<sup>val</sup> is evaluation results of `mAP(IoU=0.5)`.
74
74
- PP-YOLO_MobileNetV3 used 4 GPUs for training and mini-batch size as 32 on each GPU, if GPU number and mini-batch size is changed, learning rate and iteration times should be adjusted according [FAQ](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.0/static/docs/FAQ.md).
75
75
- PP-YOLO_MobileNetV3 inference speed is tested on Kirin 990 with 1 thread.
76
76
@@ -83,7 +83,7 @@ PP-YOLO improved performance and speed of YOLOv3 with following methods:
83
83
84
84
**Notes:**
85
85
86
-
- PP-YOLO-tiny is trained on COCO train2017 datast and evaluated on val2017 dataset,Box AP<sup>val</sup> is evaluation results of `mAP(IoU=0.5:0.95)`, Box AP<sup>val</sup> is evaluation results of `mAP(IoU=0.5)`.
86
+
- PP-YOLO-tiny is trained on COCO train2017 datast and evaluated on val2017 dataset,Box AP<sup>val</sup> is evaluation results of `mAP(IoU=0.5:0.95)`.
87
87
- PP-YOLO-tiny used 8 GPUs for training and mini-batch size as 32 on each GPU, if GPU number and mini-batch size is changed, learning rate and iteration times should be adjusted according [FAQ](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.0/static/docs/FAQ.md).
88
88
- PP-YOLO-tiny inference speed is tested on Kirin 990 with 4 threads by arm8
89
89
- we alse provide PP-YOLO-tiny post quant inference model, which can compress model to **1.3MB** with nearly no inference on inference speed and performance
0 commit comments