Warm start
In terms of Automated ML (AutoML) pipelines, hyperparameter search space can grow really quickly and an exhaustive search becomes impracticable with limited time and finite resources. You need smarter ways to perform this task, especially if you have a large dataset with a complex model working on it. If you find yourself in this kind of situation, a GridSeachCV
instances exhaustive search won't be feasible, or random parameter draws of RandomizedSearchCV
might not give you the best results given limited time.
The basic idea of warm start is to use the information gained from previous training runs to identify smarter starting points for the next training run.
For example, LogisticRegression
has a warm_start
parameter, which is set to False
by default. The following example shows you the training time the first time, and after the parameter update when it's set to False
:
from sklearn.linear_model import LogisticRegression log_reg = LogisticRegression(C=10, tol=0.00001) from sklearn...