search

naoto0804 / cross-domain-detection

Cross-Domain Weakly-Supervised Object Detection through Progressive Domain Adaptation [Inoue+, CVPR2018].
https://naoto0804.github.io/cross_domain_detection/
421 stars 77 forks source link

fine tuning in DT and PL steps? #9

Closed solapark closed 5 years ago

solapark commented 5 years ago

When I read your paper, FSD is fined-tuned both in DT and PL steps,. But, I can't find fine tuning code in your git code. Can you explain where the fine tuning part is in your code? Thank you.

naoto0804 commented 5 years ago

Thank you for your interest! Please refer here

solapark commented 5 years ago

Thank you for your reply! But when I refer there, I can't still understand that it's really fine tuning is.

In DT step, below code means you don't load pre-trained weight(=weight trained on source domain), and just 'train' network with DT images.(not fine tuning) python train_model.py --root datasets/dt_clipart/VOC2007 --root datasets/dt_clipart/VOC2012 --subset trainval --result result/dt_clipart --det_type ssd300 --data_type clipart --gpu 0 --max_iter 500 --eval_root datasets/clipart

In PL step, below code means you load pre-trained weight(=weight trained with DT images), and just 'train' network with DT images.(not fine tuning) python pseudo_label.py --root datasets/clipart --data_type clipart --det_type ssd300 --gpu 0 --load models/clipart_dt_ssd300 --result datasets/dt_pl_clipart python train_model.py --root datasets/dt_pl_clipart --subset train --result result/dt_pl_clipart --det_type ssd300 --data_type clipart --gpu 0 --load models/clipart_dt_ssd300 --eval_root datasets/clipart

As far as I now, fine tuning means freezing low level weights and updating only high level weights with new data. When I see your 'train_model.py' code, I can't find code related to fine tuning.

Am I understanding something wrong? It will be helpful if you let me know what I'm missing. Thank you.

naoto0804 commented 5 years ago

freezing low level weights was popular a few years ago, but it seems invalid nowadays, I think.