Closed domenicodigangi closed 3 years ago
should be handled automatically inheriting by torch.nn.Module and defining the parameters as torch.parameters
I have done it but then I went back to have parameters as normal torch tensors because I was having issues in extending the model from SS to Score driven. If I register a parameter as torch.nn.parameters in sequence class, and then whant to extend it to SD by defining a new class that inherits from it I need to inherit also the registeret parameters. That is an issue because they are then not straightforward to update in unrolling the SD filter
Parameters have been passed to optim functions as a vector, but that is not really needed in pytorch. Moreover organizing them in a parameters' dictionary would allow for a much easier stepwise optimization