About training loss in paper and implementation

[SuwoongHeo]

Hi,

In the paper, you said that the pretaining is conducted with the supervision of blend weights only (wgt in the below code). However, there are several losses except the skinning weight loss named diff_can and spr_wgt.

neuralgif/models/train_shape.py

Lines 156 to 160 in 40a4a96

	loss_dict['wgt'] = weight_loss
	loss_dict['diff_can'] = (diff_can +diff_can_bp)/2.0
	loss_dict['spr_wgt'] = spr_wgt
	#total_loss = weight_loss + self.loss_weight['diff_can']*loss_dict['diff_can']
	total_loss = weight_loss + self.loss_weight['diff_can']*loss_dict['diff_can'] + self.loss_weight['diff_can']**loss_dict['spr_wgt']

The definition of diff_can is obvious but spr_wgt defined below is hard to understand for me. Could you give me an explanation about what this loss means or any reference related to this loss?

neuralgif/models/train_shape.py

Lines 152 to 153 in 40a4a96

	spr_wgt = (weight_pred.abs() + 1e-12).pow(0.8).sum(1).mean() + (
	weight_smpl.abs() + 1e-12).pow(0.8).sum(1).mean()

Thanks in advance :)