在训练后期出现 nan 等情况。 #10
Description
print("means3D.shape", means3D.shape)
print("means3D",means3D.min(), means3D.max())
print("means2D.shape", means2D.shape)
print("means2D", means2D.min(), means2D.max())
print("shs.shape", shs.shape )
print("opacity.shape", opacity.shape)
print("scales.shape", scales.shape)
print("scales", scales.min(), scales.max())
print("rotations.shape", rotations.shape)
print("pc._scales.shape", pc._scaling.shape)
print("view2gaussian_precomp", view2gaussian_precomp)
我没有修改您的重要代码或者超参数，在渲染后期打印结果，我发现存在nan 值。
means3D.shape torch.Size([146342, 3])
means3D tensor(nan, device='cuda:0', grad_fn=) tensor(nan, device='cuda:0', grad_fn=)
means2D.shape torch.Size([146342, 3])
means2D tensor(0., device='cuda:0', grad_fn=) tensor(0., device='cuda:0', grad_fn=)
shs.shape torch.Size([146342, 16, 3])
opacity.shape torch.Size([146342, 1])
scales.shape torch.Size([146342, 3])
scales tensor(nan, device='cuda:0', grad_fn=) tensor(nan, device='cuda:0', grad_fn=)
rotations.shape torch.Size([146342, 4])
pc._scales.shape torch.Size([146342, 6])
view2gaussian_precomp None
