search

zhreshold / mxnet-yolo

YOLO: You only look once real-time object detector
MIT License
239 stars 88 forks source link

can't reproduce 71% mAP #20

Open tcye opened 7 years ago

tcye commented 7 years ago

I cloned your repo and follow the readme.md to train (python train.py --gpus 0,1,2,3 --epoch 0 ), but the final max mAP I can get is about 52%, did I missing something? thank you very much if you have any tricks during training!

zhreshold commented 7 years ago

You will need a pretrained model using imagenet, the one provided is in readme

Jiaolong commented 7 years ago

Hi zhreshold, I also got 52% mAP. I have downloaded the darknet19 pretrained model from the url in the readme. I think this pretrained model path is in the default arguments, right? https://github.com/zhreshold/mxnet-yolo/blob/master/train.py#L27

tcye commented 7 years ago

I have found the where is the problem, in the original code, the lr_factor is set to 1.0, just use 0.1 can get a good result, sorry for so late to report the solution

Jiaolong commented 7 years ago

@tcye Thank you! Have your reproduced the 71% mAP with lr_factor=0.1 ?

poorfriend commented 7 years ago

@zhreshold , thanks for your nice work. Is there anything which is different fron the darknet source code. Why the mAP in paper is 78.4% while we get 71.0+%.

Stinky-Tofu commented 5 years ago

I implement YOLO3 with tensorflow can only get 82% mAP without data enhancement and multi-scale training.

poorfriend commented 5 years ago

@Stinky-Tofu ， Can you share your code?

Stinky-Tofu commented 5 years ago

@Stinky-Tofu ， Can you share your code?

https://github.com/Stinky-Tofu/YOLO_V3