Can't get the real weights of the layers

[WD799]

when i use analysis_transform._layers[0].kernel and analysis_transform._layers[0].get_weights() the two functions, I get different weight. If i use analysis_transform._layers[0].kernel this function, i can get the real weights,but this kernel is read only that i can't modify them. And if i use get_weights() and set_weighs() , i get the wrong weights

[jonarchist]

Hi, this is because we use kernel DFT reparameterization by default, to improve training stability. So .get_weights() will return the DFTs of the kernels, and .kernel will return the kernel itself.

You can get the kernel using layer.kernel, and set it again using layer.kernel_parameter = RDFTParameter(kernel, name='kernel').

Or you can turn off RDFT reparameterization by instantiating the layer with the keyword kernel_parameter='variable'.