testing.CrossValidation & co.: use forkserver process start method on OS X

Author: kernc

Orange/evaluation/testing.py

@@ -1,3 +1,4 @@
+import sys
 import multiprocessing as mp
 from itertools import product
 from threading import Thread
@@ -314,6 +315,12 @@ def _mp_worker(train_data, test_data, test_i, fold_i, learner_i, learner, queue)
     return fold_i, learner_i, test_i, model, failed, predicted, probs
 
 
+def _mp_context():
+    # Workaround for locks on Macintosh
+    # https://pythonhosted.org/joblib/parallel.html#bad-interaction-of-multiprocessing-and-third-party-libraries
+    return mp.get_context('forkserver' if sys.platform == 'darwin' else None)
+
+
 class CrossValidation(Results):
     """
     K-fold cross validation.
@@ -381,7 +388,8 @@ def __init__(self, data, learners, k=10, random_state=0, store_data=False,
                 # generators are concerned. I'm stumped.
                 product(data_splits, enumerate(learners)))
 
-        with joblib.Parallel(n_jobs=n_jobs) as parallel:
+        ctx = _mp_context()
+        with joblib.Parallel(n_jobs=n_jobs, backend=ctx) as parallel:
             tasks = (joblib.delayed(_mp_worker)(*tup) for tup in args)
             thread = Thread(target=lambda: results.append(parallel(tasks)))
             thread.start()
@@ -446,7 +454,8 @@ def data_splits():
                 for (fold_i, test_i, train, test) in data_splits()
                 for (learner_i, learner) in enumerate(learners))
 
-        with joblib.Parallel(n_jobs=n_jobs) as parallel:
+        ctx = _mp_context()
+        with joblib.Parallel(n_jobs=n_jobs, backend=ctx) as parallel:
             tasks = (joblib.delayed(_mp_worker)(*tup) for tup in args)
             thread = Thread(target=lambda: results.append(parallel(tasks)))
             thread.start()
@@ -498,7 +507,8 @@ def data_splits():
                 for (fold_i, test_i, train, test) in data_splits()
                 for (learner_i, learner) in enumerate(learners))
 
-        with joblib.Parallel(n_jobs=n_jobs) as parallel:
+        ctx = _mp_context()
+        with joblib.Parallel(n_jobs=n_jobs, backend=ctx) as parallel:
             tasks = (joblib.delayed(_mp_worker)(*tup) for tup in args)
             thread = Thread(target=lambda: results.append(parallel(tasks)))
             thread.start()
@@ -565,7 +575,8 @@ def data_splits():
                 for (fold_i, test_i, train, test), (learner_i, learner) in
                 product(data_splits(), enumerate(learners)))
 
-        with joblib.Parallel(n_jobs=n_jobs) as parallel:
+        ctx = _mp_context()
+        with joblib.Parallel(n_jobs=n_jobs, backend=ctx) as parallel:
             tasks = (joblib.delayed(_mp_worker)(*tup) for tup in args)
             thread = Thread(target=lambda: results.append(parallel(tasks)))
             thread.start()
@@ -634,7 +645,8 @@ def data_splits():
                 for (fold_i, test_i, train, test) in data_splits()
                 for (learner_i, learner) in enumerate(learners))
 
-        with joblib.Parallel(n_jobs=n_jobs) as parallel:
+        ctx = _mp_context()
+        with joblib.Parallel(n_jobs=n_jobs, backend=ctx) as parallel:
             tasks = (joblib.delayed(_mp_worker)(*tup) for tup in args)
             thread = Thread(target=lambda: results.append(parallel(tasks)))
             thread.start()