Giters

Raykoooo / IAST

IAST: Instance Adaptive Self-training for Unsupervised Domain Adaptation (ECCV 2020)

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

About model selection

Songzc0 opened this issue · comments

commented

IAST/scripts/self_training_only/run_gtav2cityscapes_self_traing_only_t4.sh

Line 7 in a10bee6

python main.py --config_file config/gtav2cityscapes_t4/run_task/sl_2.yaml --resume_from ${BASE_WORK_DIR}/sl_1/epoch_2.pth --pseudo_resume_from ${BASE_WORK_DIR}/sl_1/epoch_1.pth --work_dir ${BASE_WORK_DIR}/sl_2

Thanks for your good work and the code. In the configuration files, is there any special reason to use the model after first epoch for the pseudo label generation but resume from the last model for the next round?

commented

@Songzc0 The author can answer the question better, the task is unsupervised domain adaptation, we can not know when the model is the best, and this may be a trade-off to get better results in the subsequent stages.