why ignore the grad of dir(mean) in SphericalHarmonicFunc?

[archershot]

LiteGS/gaussian_splatting/wrapper.py

Lines 317 to 334 in c34ce0e

	class SphericalHarmonicFunc(torch.autograd.Function):
	@staticmethod
	def forward(ctx,deg:int, sh_base:torch.Tensor,sh_rest:torch.Tensor, dirs:torch.Tensor):
	ctx.save_for_backward(dirs)
	ctx.degree=deg
	ctx.sh_rest_dim=sh_rest.shape[0]
	rgb=torch.ops.RasterBinning.sh2rgb_forward(deg,sh_base,sh_rest,dirs)
	return rgb
	@staticmethod
	def backward(ctx, grad_rgb):
	(dirs,)=ctx.saved_tensors
	degree=ctx.degree
	sh_rest_dim=ctx.sh_rest_dim
	sh_base_grad,sh_reset_grad=torch.ops.RasterBinning.sh2rgb_backward(degree,grad_rgb,sh_rest_dim,dirs)
	return None,sh_base_grad,sh_reset_grad,None