search

lhyfst / DSFNet

Code for DSFNet: Dual Space Fusion Network for Occlusion-Robust Dense 3D Face Alignment
58 stars 3 forks source link

training loss #5

Open xhlbetter opened 10 months ago

xhlbetter commented 10 months ago

3DMM's shape_para, exp_para, and Tm values are very large, causing a large loss in the early stages of training. How can this be addressed?

xhlbetter commented 10 months ago

@lhyfst image image image loss['shape_para_ms'] = self.rse(targets['shape_para'],shape_para_ms)1e-5 loss['shape_para_is'] = self.rse(targets['shape_para'],shape_para_is)1e-5

xhlbetter commented 9 months ago

image 想请问如何快速学到Tm的值，Tm矩阵的数据分布差异较大