search

VDIGPKU / M2Det

M2Det: A Single-Shot Object Detector based on Multi-Level Feature Pyramid Network
MIT License
1.45k stars 318 forks source link

Why adjust_learning_rate() only refers to step_lr of COCO? #113

Open yushi-okuinishi opened 4 years ago

yushi-okuinishi commented 4 years ago

https://github.com/qijiezhao/M2Det/blob/de4a6241bf22f7e7f46cb5cb1eb95615fd0a5e12/utils/core.py#L73-L80

When I tried to train with resume_epoch=300, I got the following error because max(step_lr) of COCO was 200. Traceback (most recent call last): File "train.py", line 330, in <module> lr = adjust_learning_rate(optimizer, cfg.train_cfg.gamma, epoch, step_index, iteration, epoch_size, cfg) File "/opt/program/utils/core.py", line 95, in adjust_learning_rate param_group['lr'] = lr

I think if cfg.train_cfg.step_lr.COCO[i]>=epoch was False, and lr wasn't defined.

Should we make it compatible with VOC dataset?

wangshaobobetter commented 3 years ago

change the"COCO" to "VOC"