search

facebookresearch / unbiased-teacher

PyTorch code for ICLR 2021 paper Unbiased Teacher for Semi-Supervised Object Detection
https://arxiv.org/abs/2102.09480
MIT License
415 stars 83 forks source link

using randon select date get more higher result than using COCO_supervision.txt #14

Closed happyxuwork closed 3 years ago

happyxuwork commented 3 years ago

by using the COCO_supervision.txt config when using the 10% coco as the label dataset,the results are reproducible, but when i change to the seting of random select data just as https://github.com/facebookresearch/unbiased-teacher/issues/8#issuecomment-808687654, i get more higher result:AP:36.489, under the 10% coco as the label dataset

happyxuwork commented 3 years ago

@ycliu93 do you have some Suggestions ？This is weird. when using 10% coco lable data accroding COCO_supervision.txt, can get AP :31.39, but random selecting can get AP:36.489, and i calculated the distribution of instances when using random, almost same as the COCO_supervision.txt,just as fellow diff

ycliu93 commented 3 years ago

Did you use multiple GPUs? It is likely that each GPU samples different 10% data? Could you check whether all GPUs sample the same random list?

happyxuwork commented 3 years ago

just as your said, i used multiple GPUs, and each GPU samples different 10% data, so get higher resuler. so i fiexd the seed by using np.random.seed(1), keep the each GPU samples same 10% data. thanks