Merge fb0c959 into aeca318

CHANGELOG.rst

@@ -21,6 +21,9 @@ Removed
 
 Fixed
 -----
+- Allow training of pipelines containing ``EmbeddingIntentClassifier`` in
+  a separate thread on python 3. This makes http server calls to ``/train``
+  non-blocking
 
 
 [0.13.5] - 2018-09-28

rasa_nlu/data_router.py

@@ -6,10 +6,12 @@
 import datetime
 import io
 import logging
+import multiprocessing
 import os
 from concurrent.futures import ProcessPoolExecutor as ProcessPool
 from typing import Text, Dict, Any, Optional, List
 
+import six
 from builtins import object
 from twisted.internet import reactor
 from twisted.internet.defer import Deferred
@@ -107,6 +109,13 @@ def __init__(self,
             self.component_builder = ComponentBuilder(use_cache=True)
 
         self.project_store = self._create_project_store(project_dir)
+
+        if six.PY3:
+            # tensorflow sessions are not fork-safe,
+            # and training processes have to be spawned instead of forked.
+            # See https://github.com/tensorflow/tensorflow/issues/5448#issuecomment-258934405
+            multiprocessing.set_start_method('spawn', force=True)
+
         self.pool = ProcessPool(self._training_processes)
 
     def __del__(self):
@@ -358,9 +367,9 @@ def training_errback(failure):
         self._current_training_processes += 1
         self.project_store[project].current_training_processes += 1
 
-        # tensorflow training is not executed in a separate thread, as this may
-        # cause training to freeze
-        if self._tf_in_pipeline(train_config):
+        # tensorflow training is not executed in a separate thread on python 2,
+        # as this may cause training to freeze
+        if six.PY2 and self._tf_in_pipeline(train_config):
             try:
                 logger.warning("Training a pipeline with a tensorflow "
                                "component. This blocks the server during "