/rfcn/../lib/bbox/bbox_transform.py:82: RuntimeWarning: divide by zero encountered in divide targets_dw = np.log(gt_widths / ex_widths) /rfcn/../lib/bbox/bbox_transform.py:114: RuntimeWarning: overflow encountered in exp pred_w = np.exp(dw) * widths[:, np.newaxis] /rfcn/../lib/bbox/bbox_transform.py:115: RuntimeWarning: overflow encountered in exp pred_h = np.exp(dh) * heights[:, np.newaxis]

[hdjsjyl]

Dear author, thanks for you code. When I change it to my own dataset, I met this problem which makes the RPNL1Loss very big such as 22719 or nan. Do you know which part produces this problem?

error: /rfcn/../lib/bbox/bbox_transform.py:82: RuntimeWarning: divide by zero encountered in divide targets_dw = np.log(gt_widths / ex_widths) /rfcn/../lib/bbox/bbox_transform.py:114: RuntimeWarning: overflow encountered in exp pred_w = np.exp(dw) * widths[:, np.newaxis] /rfcn/../lib/bbox/bbox_transform.py:115: RuntimeWarning: overflow encountered in exp pred_h = np.exp(dh) * heights[:, np.newaxis]

[Kongsea]

I have met the same error. However, when training the same dataset using faster-rcnn, everything is OK. And it's also OK when training the same dataset using FPN in Detectron. Is there anybody to give some advice? Thank you.

[Kongsea]

Besides, I have changed the learning rate to 1e-5 or even small, but it still didn't work.

[yhs95]

After I alter lr to lr=0.0025,this problem is solved.