Continue training

[romanovzky]

Hi,

I was testing whether I could fit a SymbolicRegressor up to, say, 1000 generations, see the Pareto front, and then continue training for another 1000 generations. However, it seems that if I do

reg = SymbolicRegressor()
reg.fit(X_train, y_train)

play with reg and then

reg.fit(X_train, y_train)

again, the reg object is the same as before the second fit call. Now, given that reg has a Pareto front, wouldn't I be able to continue fitting "a la" online learning/batch/partial fit way? I'm trying to "brute force" a work-around for the lack of callbacks (see #18).

Cheers

[foolnotion]

Hi, this looks like an easy improvement, but it will still require some changes to the C++ library. Currently, each call to fit initializes a new C++ algorithm, runs it, and keep some stats and results (like the pareto front) from it. But when fit is done the C++ object doesn't exist anymore. However, it should be easy enough to implement a kind of warm start mechanism.

[romanovzky]

Yes, a warm start mechanism would be super useful! Already thinking about the possibility of using things like hyperband, which ideally need warm start mechanisms.