gaobb / DCFS

[NeurIPS 2022] Decoupling Classifier for Boosting Few-shot Object Detection and Instance Segmentation
https://csgaobb.github.io/Projects/DCFS
MIT License
29 stars 2 forks source link

脚本文件里好像有错,可以确认下吗? #6

Closed lixi92 closed 9 months ago

lixi92 commented 9 months ago

----------------------------- Model Preparation --------------------------------- python3 tools/model_surgery.py --dataset voc --method randinit \ --src-path ${SAVEDIR}/dcfsdet${NET}_base${SPLIT_ID}/model_final.pth \ --save-dir ${SAVEDIR}/dcfsdet${NET}_base${SPLIT_ID} BASE_WEIGHT=${SAVEDIR}/dcfsdet${NET}_base${SPLIT_ID}/model_reset_surgery.pth

------------------------------ Novel Fine-tuning ------------------------------- --> 2. TFA-like, i.e. run seed0~9 10 times for gFSOD on voc classloss="DC" # "CE"

for seed in 0 1 2 3 4 5 6 7 8 9

第二次fine-tune使用的预训练权重应该是第一次fine-tune结果目录下的权重?

gaobb commented 9 months ago

非常感谢指出该问题。你是对的,应该使用base model weight。

之所以出现这个问题,是当时训练完我再次评测,代码发布时忘记删掉当前模型的权重了。现在已删除并更新。