the result of solov2 model after distiller is reduced about fgd HOT 6 CLOSED

It seems your teacher doesn't load the weight exactly. Besides, the inheriting strategy seems wrong either.

from fgd.

You mean the teacher model didn't load successfully from the beginning, what's wrong in inheriting strategy?
The teacher model was trained by myself before distiller , the config file is below:
solov2_r101_fpn_3x_c8_twoc.txt
I change the mobileNetV2 backbone channel from 2048 to 320 in student model in order to slim the model .

from fgd.

How do you write the code in this part?

from fgd.

The code in mmdet/distillation/distillers/detection_distiller.py, I am not change.

from fgd.

You use mmdet==2.24, so you should add self.student.init_weights() in mmdet/distillation/distiller.py. Besides, you don't need model.init_weights() in tools/train.py for distiller.

from fgd.

OK, I do use mmdet==2.24. I have revised these two lines and train agin. Thank you so much again.

from fgd.