search

Megvii-BaseDetection / YOLOX

YOLOX is a high-performance anchor-free YOLO, exceeding yolov3~v5 with MegEngine, ONNX, TensorRT, ncnn, and OpenVINO supported. Documentation: https://yolox.readthedocs.io/
Apache License 2.0
9.39k stars 2.2k forks source link

use ciou training #986

Closed aidemHJC closed 2 years ago

aidemHJC commented 2 years ago

Dear sir could pls help me, I add Diou and Ciou in losses.py, and Diou works well, the total loss decrease with epoch increase, BUT Ciou, the total loss decease at beginning epochs, then increase, especially the conf_loss attach file is losses.py and training log, my code downloaded at 2021/10/2 thanks, BR code and log.zip 8

aidemHJC commented 2 years ago

my code downloaded at 2021/10/28

scsolaris commented 2 years ago

@aidemHJC using yolov5 version of mosaic augmentation could increase the training stability

aidemHJC commented 2 years ago

@scsolaris do you mean that using ciou with my code and yolov5 version of mosaic augmentation the same time can increase the training stability?

scsolaris commented 2 years ago

@scsolaris do you mean that using ciou with my code and yolov5 version of mosaic augmentation the same time can increase the training stability?

@aidemHJC yes. yolov5 version of mosaic augmentation filters out some low quality boxes.

vaerdu commented 2 years ago

@scsolaris do you mean that using ciou with my code and yolov5 version of mosaic augmentation the same time can increase the training stability?

@aidemHJC yes. yolov5 version of mosaic augmentation filters out some low quality boxes.

interesting! do you have open source code?

scsolaris commented 2 years ago

@vaerdu yolox has used the yolov5 version of mosaic augmentation. please refer to #765

aidemHJC commented 2 years ago

@scsolaris thanks for your reply