如何将 sklearn 管道转换为 pyspark 管道？

Question

We have a machine learning classifier model that we have trained with a pandas dataframe and a standard sklearn pipeline (StandardScaler, RandomForestClassifier, GridSearchCV etc). We are working on Databricks and would like to scale up this pipeline to a large dataset using the parallel computation spark offers.

What is the quickest way to convert our sklearn pipeline into something that computes in parallel ? (We can easily switch between pandas and spark DFs as required.)

For context, our options seem to be:

1. Rewrite the pipeline using MLLib (time-consuming)
2. Use a sklearn-spark bridging library

On option 2, Spark-Sklearn seems to be deprecated , but Databricks instead recommends that we use joblibspark. However, this raises an exception on Databricks:

from sklearn import svm, datasets
from sklearn.model_selection import GridSearchCV
from joblibspark import register_spark
from sklearn.utils import parallel_backend
register_spark() # register spark backend

iris = datasets.load_iris()
parameters = {'kernel':('linear', 'rbf'), 'C':[1, 10]}
svr = svm.SVC(gamma='auto')

clf = GridSearchCV(svr, parameters, cv=5)
with parallel_backend('spark', n_jobs=3):
    clf.fit(iris.data, iris.target)

raises

py4j.security.Py4JSecurityException: Method public int org.apache.spark.SparkContext.maxNumConcurrentTasks() is not whitelisted on class class org.apache.spark.SparkContext

Answer 1

According to the Databricks instructions ( here and here ), the necessary requirements are:

• Python 3.6+
• pyspark>=2.4
• scikit-learn>=0.21
• joblib>=0.14

I cannot reproduce your issue in a community Databricks cluster running Python 3.7.5, Spark 3.0.0, scikit-learn 0.22.1, and joblib 0.14.1:

import sys
import sklearn
import joblib

spark.version
# '3.0.0'

sys.version
# '3.7.5 (default, Nov  7 2019, 10:50:52) \n[GCC 8.3.0]'

sklearn.__version__
# '0.22.1'

joblib.__version__
# '0.14.1'

With the above settings, your code snippet runs smoothly, and produces indeed a classifier clf as:

GridSearchCV(cv=5, error_score=nan,
             estimator=SVC(C=1.0, break_ties=False, cache_size=200,
                           class_weight=None, coef0=0.0,
                           decision_function_shape='ovr', degree=3,
                           gamma='auto', kernel='rbf', max_iter=-1,
                           probability=False, random_state=None, shrinking=True,
                           tol=0.001, verbose=False),
             iid='deprecated', n_jobs=None,
             param_grid={'C': [1, 10], 'kernel': ('linear', 'rbf')},
             pre_dispatch='2*n_jobs', refit=True, return_train_score=False,
             scoring=None, verbose=0)

as does the alternative example from here :

from sklearn.utils import parallel_backend
from sklearn.model_selection import cross_val_score
from sklearn import datasets
from sklearn import svm
from joblibspark import register_spark

register_spark() # register spark backend

iris = datasets.load_iris()
clf = svm.SVC(kernel='linear', C=1)
with parallel_backend('spark', n_jobs=3):
  scores = cross_val_score(clf, iris.data, iris.target, cv=5)

print(scores)

giving

[0.96666667 1.         0.96666667 0.96666667 1.        ]

Answer 2

Thanks to desertnaut for the response - this answer should be correct for a standard Spark / Databricks setup, so have accepted it, given the wording of my question / potential usefulness for other readers

Contributing a separate "answer" having discovered what the issue was in our case: Databricks support advised that the issue in our case was due to our using a special type of cluster (High Concurrency with credentials passthrough enabled, on AWS). grid.fit() was not whitelisted for this type of cluster, and Databricks advised that they would need to raise it with their engineering team to whitelist it.