efficientdet-pytorch
efficientdet-pytorch copied to clipboard
coderhss
Pre-trained or checkpoint?
First of all, I am very impressed and grateful. It seems that you already ran the training process. It will be amazing if you share the checkpoint/pre-trained weight. @coderhss
If you did evaluation of the final model, it will be amazing to put the performance. So that we can compare it to the original paper.
This code has not yet fully reproduced the results of the paper, and I will give the results of my current training today.
Thank you very much for sharing, I just want to inform you that EfficientDet-D0/D1/D3/D4 uses RetinaNet preprocessing while EfficientDet-D5/D6/D7 uses auto-augmentation preprocessing. maybe that is why your result cannot achieve the paper results.
HaHa , I saw so many guys who wants to reproduce the result these days , but all of them falied .
HaHa , I saw so many guys who wants to reproduce the result these days , but all of them falied .
yes, that is correct. What is so funny about that? those "guys" are noble for trying.
HaHa , I saw so many guys who wants to reproduce the result these days , but all of them falied .
yes, that is correct. What is so funny about that? those "guys" are noble for trying.
Because those technique described in the paper is not all the tricks they used . But everyone trust the paper without hesitate . Just because it is published by Google.
@coderhss Thank you for the great repo! I'm wondering if you have tried D1-D3 yet? If you did, can you share the mAP for D1-D3 models? Thank you so much!
