Currently skgrf executes predictions by passing the serialized forest dict. The prediction functions deserialize the forest on every prediction, which is expensive. This PR modifies the prediction functions so that they accept pointers to already deserialized forest objects.
In order to do this, we create a Cython GRFForest class which serves as a container for the forest pointer. We also add an _ensure_ptr method, which creates this forest wrapper for a trained estimator if it does not already exist.
Since this wrapped pointer cannot be serialized, we also override the __getstate__ and __setstate__ methods to ensure we can pickle and unpickle each class. On unpickling, we also call _ensure_ptr so that the first call to predict after unpickling is not cold-started.
The result of this is that predict calls are faster. Here is a simple comparison with sample data from my experimenting with this:
import timeit
from skgrf.ensemble import GRFRegressor
from sklearn.datasets import load_boston
boston_X, boston_y = load_boston(return_X_y=True)
gfr = GRFRegressor()
gfr.fit(boston_X, boston_y)
ptr = timeit.timeit(lambda : gfr.predict_ptr(boston_X), number=100)
ser = timeit.timeit(lambda : gfr.predict(boston_X), number=100)
print(ptr)
0.252420819000001
print(ser)
1.2568888119999997
Changes
Add a GRFForest wrapper around a forest pointer in Cython, also modify all Cython predict functions to accept this object in lieu of the serialized forest dict.
Add _ensure_ptr() method which creates GRFForest objects on trained estimators if not already existing
Override pickle serde methods to ensure we can serde each class, because GRFForest objects are not serializable
Fix the instance variable of the GRFBoostedRegressor from grf_forest_ to boosted_forests_ in docs
Rename GRFValidationMixin to GRFMixin (it's not just validation now)
Currently skgrf executes predictions by passing the serialized forest dict. The prediction functions deserialize the forest on every prediction, which is expensive. This PR modifies the prediction functions so that they accept pointers to already deserialized forest objects.
In order to do this, we create a Cython
GRFForestclass which serves as a container for the forest pointer. We also add an_ensure_ptrmethod, which creates this forest wrapper for a trained estimator if it does not already exist.Since this wrapped pointer cannot be serialized, we also override the
__getstate__and__setstate__methods to ensure we can pickle and unpickle each class. On unpickling, we also call_ensure_ptrso that the first call to predict after unpickling is not cold-started.The result of this is that predict calls are faster. Here is a simple comparison with sample data from my experimenting with this:
Changes
GRFForestwrapper around a forest pointer in Cython, also modify all Cython predict functions to accept this object in lieu of the serialized forest dict._ensure_ptr()method which createsGRFForestobjects on trained estimators if not already existingGRFForestobjects are not serializableGRFBoostedRegressorfromgrf_forest_toboosted_forests_in docsGRFValidationMixintoGRFMixin(it's not just validation now)