[###updated for new question] rcnn_loss don't decline when training on my custum dataset

[xzswater]

As shown on the fig, ### rcnn_loss don't decline sharply; on the contrary, rpn_loss worked well.
So, what are the possible reasons ? Thank you for your suggestion.

[sshaoshuai]

How about the final mAP? Sometimes the rcnn loss curve could not well indicate the training process of the network. Also, the rcnn_fg_num seems too low. Maybe you should check the quality of RPN proposals. And also you could modify the RCNN.TARGET_CONFIG.[REG_FG_THRESH, CLS_FG_THRESH, CLS_BG_THRESH] according to your dataset or classes.