some questions about loce

[sanmulab]

Hi author, thank you for your excellent work! I have some questions:

1. Can MFS be applied to other loss functions, such as EQLv2 and Seesaw Loss, have you done these experiments?
2. The decoupling training only trains 6 epochs. Will this be too little? Will training more epochs still improve AP?

[fcjian]

1. MFS is independent with the loss function, and can be applied to other loss functions theoretically, but we didn't conduct the experiments on EQLv2 and Seesaw Loss.
2. We fine-tune the model for 6 epochs following the previous works such as BALMS, and we didn't fine-tune the model for more epochs.

[sanmulab]

Thanks for your reply!

[sanmulab]

Hi! I have one last question: how to use MFS alone for training, and how should I modify the config or code? In addition to replacing the EBL loss function, do I need to modify LoceRoIHead?

[fcjian]

Hi! I have one last question: how to use MFS alone for training, and how should I modify the config or code? In addition to replacing the EBL loss function, do I need to modify LoceRoIHead?

For the config, you should modify the loss configure:

• LOCE/configs/loce/loce_mask_rcnn_r50_fpn_normed_mask_mstrain_2x_lvis_v1.py

  Lines 8 to 10 in 0a0194d

  	loss_cls=dict(
  	type='EquilibriumLoss', use_sigmoid=False, loss_weight=1.0),
  	)

For the code, you should modify the loss inputs:

• LOCE/mmdet/models/roi_heads/bbox_heads/loce_convfc_bbox_head.py

  Lines 214 to 220 in 0a0194d

  	losses['loss_cls'] = self.loss_cls(
  	cls_score,
  	labels,
  	label_weights,
  	mean_score,
  	avg_factor=avg_factor,
  	reduction_override=reduction_override)

[sanmulab]

Sorry, I still don't know which loss parameters need to be modified. If I just use CrossEntropyLoss+MFS for training, then which parameters should I modify in loce_convfc_bbox_head.py:
losses['loss_cls'] = self.loss_cls(
cls_score,
labels,
label_weights,
mean_score,
avg_factor=avg_factor,
reduction_override=reduction_override)

[fcjian]

You should remove mean_score.

And you can refer to the code:

• LOCE/mmdet/models/roi_heads/bbox_heads/bbox_head.py

  Lines 232 to 237 in 0a0194d

  	losses['loss_cls'] = self.loss_cls(
  	cls_score,
  	labels,
  	label_weights,
  	avg_factor=avg_factor,
  	reduction_override=reduction_override)

[sanmulab]

Thank you!