scaling sklearn RandomForestClassifier for RandomizedSearchCV

Question

I'm training a sklearn.ensemble.RandomForestClassifier() on a single cluster node that has 28 CPUs and ~190GB RAM. Training this classifier alone runs quite fast, uses all cores on the machine and uses ~93GB RAM:

x_train,x_test,y_train,y_test=sklearn.model_selection.train_test_split(x,y,test_size=.25,random_state=0)

clf=sklearn.ensemble.RandomForestClassifier(n_estimators=100,
                                            random_state=0,
                                            n_jobs=-1,
                                            max_depth=10,
                                            class_weight='balanced',
                                            warm_start=False,
                                            verbose=2)
clf.fit(x_train,y_train)

with output:

[Parallel(n_jobs=-1)]: Done  88 out of 100 | elapsed:  1.9min remaining:   15.2s
[Parallel(n_jobs=-1)]: Done 100 out of 100 | elapsed:  2.0min finished
CPU times: user 43min 10s, sys: 1min 33s, total: 44min 44s
Wall time: 2min 20s

However, this particular model seems not optimal, having performance ~80% correct. So I want to optimize hyperparameters for the model using sklearn.model_selection.RandomizedSearchCV(). So I setup the search like so:

rfc = sklearn.ensemble.RandomForestClassifier()
rf_random = sklearn.model_selection.RandomizedSearchCV(estimator=rfc, 
                                                       param_distributions=random_grid, 
                                                       n_iter=100, 
                                                       cv=3, 
                                                       verbose=2, 
                                                       random_state=0, 
                                                       n_jobs=2, 
                                                       pre_dispatch=1)
rf_random.fit(x, y)

But I cannot find setting for n_jobs and pre_dispatch that uses the hardware effectively. Here are some example runs and the results:

n_jobs   pre_dispatch    Result
===========================================================================
default       default    Utilizes all cores but Job killed - out of memory
    -1              1    Job killed - out of memory
    12              1    Job killed - out of memory
     3              1    Job killed - out of memory
     2              1    Job runs, but only utilizes 2 cores, takes >230min (wall clock) per model

How can I get the performance that I see when training a standalone RandomForestClassifier when running a hyperparameter search? And how is the standalone version parallelizing such that it does not create copies of my large dataset like with the grid search?

---

EDIT: The following combination of parameters effectively used all cores for training each individual RandomForestClassifier without parallelizing the hyperparameter search itself or blowing up the RAM usage.

model = sklearn.ensemble.RandomForestClassifier(n_jobs=-1, verbose=1)
search = sklearn.model_selection.RandomizedSearchCV(estimator=model, 
                                                    param_distributions=random_grid, 
                                                    n_iter=10, 
                                                    cv=3, 
                                                    verbose=10, 
                                                    random_state=0,
                                                    n_jobs=1,
                                                    pre_dispatch=1)
with joblib.parallel_backend('threading'):
    search.fit(x, y)

Answer 1

If the single classifier training saturates all your cores, then there is nothing to gain by parallelizing the gridsearch also. Set n_jobs=1 for gridsearch, and keep n_jobs=-1 for the classifier. This should avoid the out of memory condition.