<11% mAP with pretrained YOLO after tiny amount of training

eriklindernoren / PyTorch-YOLOv3

Minimal PyTorch implementation of YOLOv3

GNU General Public License v3.0

7.31k stars 2.63k forks source link

<11% mAP with pretrained YOLO after tiny amount of training #278

Open half-potato opened 5 years ago

half-potato commented 5 years ago

Only 200 iterations of training on COCO completely destroys the accuracy of the pretrained YOLOv3 net. I am using an unchanged repo with the COCO dataset downloaded using the script and version 1.1.0 of pytorch.

Why does this happen? Do I actually need to train it from scratch completely?

fadi212 commented 5 years ago

Yes. It has updated the weights file :D Now just remove the yolov3.weights file and replace it with the new file. I think this is the only explanation from me