Looking at this GitHub issue I am afraid that's not possible.
In my experience working with the library, I have never worked on a use case that required such modification, I mostly change the values for the arguments in the constructor of the estimators.
I think that if you want to customise an algorithm (e.g., logistic regression) more, then you'd have to implement it yourself. On the other hand, I've found this StackOverflow question that specifies how you can implement a custom loss function for
sklearn estimators, so maybe there's a workaround for it, even though it seems that the library was not designed to be used like that.
LogisticRegressiondoes support the L1 and L2 penalties, but otherwise I agree with @StefanPopov's answer that it's not generally possible. $\endgroup$