Some questions about the trainable parameters?

[Sundrops]

Thanks for your great code!
self.centered_gradient_kernels = self.get_centered_gradient_kernel().train(False)
When I remove the 'train(False)', I found that the paramters of get_centered_gradient_kernel cannot be trained and it retains unchanged.I guess that maybe some ops are non-differentiable. And as the paper said "We also propose to relax the convolutional filters in Eq. (7)-(9). The original convolutions are used to derive the (numerical) gradients and divergences", I wonder why this implement and the original implement don't set this paramter trainable. If so, the trainable paramter is only u0. And the result of trained model is similar to the untrained model's.