Why self.model_state_dict changes during valuation when using pytorchclassfier

[lwx0420]

In my understanding, self.model_state_dict stores the original parameters of the model, but i find it changes during valuation. And i can't find out why this happens.
I find this question when I save the data valuation result and original model parameters so that i can run data_removal() later. Then i found the self.model_state_dict before and after dv.estimate() is different.