DETRHead.cls_out_channels and FocalLoss.use_sigmoid

[ShunchiZhang]

Hi, would you mind illustrate why self.cls_out_channels = num_classes + 1 when not using sigmoid?

mmdetection/mmdet/models/dense_heads/detr_head.py

Lines 134 to 137 in 9d3e162

	if self.loss_cls.use_sigmoid:
	self.cls_out_channels = num_classes
	else:
	self.cls_out_channels = num_classes + 1

Besides, why use_sigmoid = False is not supported in focal loss class? How could I use focal loss with softmax?

mmdetection/mmdet/models/losses/focal_loss.py

Line 188 in 9d3e162

assert use_sigmoid is True, 'Only sigmoid focal loss supported now.'

Thank you so much!

[zachcoleman]

In general, when use_sigmoid=True you are doing the binary cross-entropy per each class individually. Since every probability from each channel is [0,1] you do not need a background class. When use_sigmoid=False you are doing softmax typically and the probabilities across each class sum to 1. Thus, you need a background class for not any class.

As far as your other question. It appears it just hasn't been implemented yet.