Hi zhenghao,
the train/eval mode of the fcos model is written in your code [path/model/fcos.py], however, in the line 26
'super().train(mode=True)', the model mode is fixed as training mode.
That leads to some unexpected inference results when doing inference by set model to eval(). So, I suggest
modify 'super().train(mode=True)' to 'super().train(mode=mode)'.
Otherwise, the dropout/bn is not locked in the inference procedure. At the same time, the dropout/bn status
can be affected by the [path/model/config.py] configurations: freeze_stage1/freeze_bn
Hi zhenghao, the train/eval mode of the fcos model is written in your code [path/model/fcos.py], however, in the line 26 'super().train(mode=True)', the model mode is fixed as training mode. That leads to some unexpected inference results when doing inference by set model to eval(). So, I suggest modify 'super().train(mode=True)' to 'super().train(mode=mode)'. Otherwise, the dropout/bn is not locked in the inference procedure. At the same time, the dropout/bn status can be affected by the [path/model/config.py] configurations: freeze_stage1/freeze_bn