training loss

[xhlbetter]

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]

@lhyfst



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]

想请问如何快速学到Tm的值，Tm矩阵的数据分布差异较大