Giters

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/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

ImageNet Pretraining

ocetintas opened this issue · comments

commented

Hi,
I am looking for ImageNet pretrained weights of the YOLOX backbone. I am specifically interested in the largest model YOLOX-x. In a couple of other issues I've seen that nano version can be trained from scratch for the same performance, but there is no information about the x version.

Overall, does ImageNet training not help at all? If there are no pretrained weights but ImageNet pretraining does help, I will invest time into training the model myself. So, I'd really appreciate your help @FateScript

Thank you in advance

commented

Hi, @ocetintas If you train your network with enough epoch/iter/data, the performance will be nearly the same. The effect of imagenet pretraining is making your network achieve a modest performance with less epoch/iter/data.
However, to achieve a best performance, imagenet pretraining couldn't help much.