Improve random state handling

CHANGELOG.md

@@ -5,6 +5,11 @@ All notable changes to this project will be documented in this file.
 ### Changed
 - Re-score ambiguous `DeterministicIntentParser` results based on slots [#791](https://github.com/snipsco/snips-nlu/pull/791)
 - Accept ambiguous results from `DeterministicIntentParser` when confidence score is above 0.5 [#797](https://github.com/snipsco/snips-nlu/pull/797)
+- Moved the NLU random state from the config to the shared resources [#801](https://github.com/snipsco/snips-nlu/pull/801)
+- Bumped `scikit-learn` to `>=0.21,<0.22` for `python>=3.5` and `>=0.20<0.21` for `python<3.5` [#801](https://github.com/snipsco/snips-nlu/pull/801) 
+
+### Fixed
+- Fixed a couple of bugs in the data augmentation which were making the NLU training non-deterministic [#801](https://github.com/snipsco/snips-nlu/pull/801)
 
 ## [0.19.6]
 ### Fixed

docs/source/tutorial.rst

@@ -174,6 +174,26 @@ the dataset we generated earlier:
 
     engine.fit(dataset)
 
+Note that, by default, training of the NLU engine is non-deterministic:
+training and testing multiple times on the same data may produce different
+outputs.
+
+Reproducible trainings can be achieved by passing a **random seed** to the
+engine:
+
+.. code-block:: python
+
+    seed = 42
+    engine = SnipsNLUEngine(config=CONFIG_EN, random_state=seed)
+    engine.fit(dataset)
+
+
+.. note::
+
+    Due to a ``scikit-learn`` bug fixed in version ``0.21`` we can't guarantee
+    any deterministic behavior if you're using a Python version ``<3.5`` since
+    ``scikit-learn>=0.21`` is only available starting from Python ``>=3.5``
+
 
 Parsing
 -------

setup.py

@@ -21,18 +21,19 @@
     "future>=0.16,<0.17",
     "numpy>=1.15,<1.16",
     "scipy>=1.0,<2.0",
-    "scikit-learn>=0.19,<0.20",
+    "scikit-learn>=0.21.1,<0.22; python_version>='3.5'",
+    "scikit-learn>=0.20,<0.21; python_version<'3.5'",
     "sklearn-crfsuite>=0.3.6,<0.4",
     "semantic_version>=2.6,<3.0",
     "snips-nlu-utils>=0.8,<0.9",
     "snips-nlu-parsers>=0.2,<0.3",
     "num2words>=0.5.6,<0.6",
     "plac>=0.9.6,<1.0",
     "requests>=2.0,<3.0",
-    "pathlib==1.0.1; python_version < '3.4'",
+    "pathlib==1.0.1; python_version<'3.4'",
     "pyaml>=17,<18",
     "deprecation>=2,<3",
-    "funcsigs>=1.0,<2.0; python_version < '3.4'"
+    "funcsigs>=1.0,<2.0; python_version<'3.4'"
 ]
 
 extras_require = {

snips_nlu/constants.py

@@ -46,6 +46,7 @@
 BUILTIN_ENTITY_PARSER = "builtin_entity_parser"
 CUSTOM_ENTITY_PARSER = "custom_entity_parser"
 MATCHING_STRICTNESS = "matching_strictness"
+RANDOM_STATE = "random_state"
 
 # resources
 RESOURCES = "resources"

snips_nlu/data_augmentation.py

@@ -69,10 +69,10 @@ def get_entities_iterators(intent_entities, language,
                            add_builtin_entities_examples, random_state):
     entities_its = dict()
     for entity_name, entity in iteritems(intent_entities):
-        utterance_values = random_state.permutation(list(entity[UTTERANCES]))
+        utterance_values = random_state.permutation(sorted(entity[UTTERANCES]))
         if add_builtin_entities_examples and is_builtin_entity(entity_name):
-            entity_examples = get_builtin_entity_examples(entity_name,
-                                                          language)
+            entity_examples = get_builtin_entity_examples(
+                entity_name, language)
             # Builtin entity examples must be kept first in the iterator to
             # ensure that they are used when augmenting data
             iterator_values = entity_examples + list(utterance_values)

snips_nlu/default_configs/config_de.py

@@ -111,8 +111,7 @@
                     "min_utterances": 200,
                     "capitalization_ratio": 0.2,
                     "add_builtin_entities_examples": True
-                },
-                "random_seed": None
+                }
             },
             "intent_classifier_config": {
                 "unit_name": "log_reg_intent_classifier",
@@ -140,8 +139,7 @@
                         "unknown_words_replacement_string": None,
                         "keep_order": True
                     }
-                },
-                "random_seed": None
+                }
             }
         }
     ]

snips_nlu/default_configs/config_en.py

@@ -97,8 +97,7 @@
                     "min_utterances": 200,
                     "capitalization_ratio": 0.2,
                     "add_builtin_entities_examples": True
-                },
-                "random_seed": None
+                }
             },
             "intent_classifier_config": {
                 "unit_name": "log_reg_intent_classifier",
@@ -126,8 +125,7 @@
                         "unknown_words_replacement_string": None,
                         "keep_order": True
                     }
-                },
-                "random_seed": None
+                }
             }
         }
     ]

snips_nlu/default_configs/config_es.py

@@ -90,7 +90,7 @@
                     "capitalization_ratio": 0.2,
                     "add_builtin_entities_examples": True
                 },
-                "random_seed": None
+
             },
             "intent_classifier_config": {
                 "unit_name": "log_reg_intent_classifier",
@@ -118,8 +118,7 @@
                         "unknown_words_replacement_string": None,
                         "keep_order": True
                     }
-                },
-                "random_seed": None
+                }
             }
         }
     ]

snips_nlu/default_configs/config_fr.py

@@ -89,8 +89,7 @@
                     "min_utterances": 200,
                     "capitalization_ratio": 0.2,
                     "add_builtin_entities_examples": True
-                },
-                "random_seed": None
+                }
             },
             "intent_classifier_config": {
                 "unit_name": "log_reg_intent_classifier",
@@ -118,8 +117,7 @@
                         "unknown_words_replacement_string": None,
                         "keep_order": True
                     }
-                },
-                "random_seed": None
+                }
             }
         }
     ]

snips_nlu/default_configs/config_it.py

@@ -89,8 +89,7 @@
                     "min_utterances": 200,
                     "capitalization_ratio": 0.2,
                     "add_builtin_entities_examples": True
-                },
-                "random_seed": None
+                }
             },
             "intent_classifier_config": {
                 "unit_name": "log_reg_intent_classifier",
@@ -118,8 +117,7 @@
                         "unknown_words_replacement_string": None,
                         "keep_order": True
                     }
-                },
-                "random_seed": None
+                }
             }
         }
     ]

snips_nlu/default_configs/config_ja.py

@@ -116,7 +116,7 @@
                     "capitalization_ratio": 0.2,
                     "add_builtin_entities_examples": True
                 },
-                "random_seed": None
+
             },
             "intent_classifier_config": {
                 "unit_name": "log_reg_intent_classifier",
@@ -144,8 +144,7 @@
                         "unknown_words_replacement_string": None,
                         "keep_order": True
                     }
-                },
-                "random_seed": None
+                }
             }
         }
     ]

snips_nlu/default_configs/config_ko.py

@@ -107,8 +107,7 @@
                     "min_utterances": 200,
                     "capitalization_ratio": 0.2,
                     "add_builtin_entities_examples": True
-                },
-                "random_seed": None
+                }
             },
             "intent_classifier_config": {
                 "unit_name": "log_reg_intent_classifier",
@@ -136,8 +135,7 @@
                         "unknown_words_replacement_string": None,
                         "keep_order": True
                     }
-                },
-                "random_seed": None
+                }
             }
         }
     ]

snips_nlu/intent_classifier/featurizer.py

@@ -109,7 +109,9 @@ def _fit_transform_tfidf_vectorizer(self, x, y, dataset):
             config=self.config.tfidf_vectorizer_config,
             builtin_entity_parser=self.builtin_entity_parser,
             custom_entity_parser=self.custom_entity_parser,
-            resources=self.resources)
+            resources=self.resources,
+            random_state=self.random_state,
+        )
         x_tfidf = self.tfidf_vectorizer.fit_transform(x, dataset)
 
         if not self.tfidf_vectorizer.vocabulary:
@@ -139,7 +141,9 @@ def _fit_cooccurrence_vectorizer(self, x, classes, none_class, dataset):
             config=self.config.cooccurrence_vectorizer_config,
             builtin_entity_parser=self.builtin_entity_parser,
             custom_entity_parser=self.custom_entity_parser,
-            resources=self.resources)
+            resources=self.resources,
+            random_state=self.random_state,
+        )
         x_cooccurrence = self.cooccurrence_vectorizer.fit(
             non_null_x, dataset).transform(x)
         if not self.cooccurrence_vectorizer.word_pairs:

snips_nlu/intent_classifier/log_reg_classifier.py

@@ -10,11 +10,10 @@
 
 from snips_nlu.common.log_utils import DifferedLoggingMessage, log_elapsed_time
 from snips_nlu.common.utils import (
-    check_persisted_path, check_random_state,
-    fitted_required, json_string)
+    check_persisted_path, fitted_required, json_string)
 from snips_nlu.constants import LANGUAGE, RES_PROBA
 from snips_nlu.dataset import validate_and_format_dataset
-from snips_nlu.exceptions import _EmptyDatasetUtterancesError, LoadingError
+from snips_nlu.exceptions import LoadingError, _EmptyDatasetUtterancesError
 from snips_nlu.intent_classifier.featurizer import Featurizer
 from snips_nlu.intent_classifier.intent_classifier import IntentClassifier
 from snips_nlu.intent_classifier.log_reg_classifier_utils import (
@@ -24,11 +23,20 @@
 
 logger = logging.getLogger(__name__)
 
+# We set tol to 1e-3 to silence the following warning with Python 2 (
+# scikit-learn 0.20):
+#
+# FutureWarning: max_iter and tol parameters have been added in SGDClassifier
+# in 0.19. If max_iter is set but tol is left unset, the default value for tol
+# in 0.19 and 0.20 will be None (which is equivalent to -infinity, so it has no
+# effect) but will change in 0.21 to 1e-3. Specify tol to silence this warning.
+
 LOG_REG_ARGS = {
     "loss": "log",
     "penalty": "l2",
     "class_weight": "balanced",
-    "max_iter": 5,
+    "max_iter": 1000,
+    "tol": 1e-3,
     "n_jobs": -1
 }
 
@@ -66,12 +74,11 @@ def fit(self, dataset):
         self.fit_builtin_entity_parser_if_needed(dataset)
         self.fit_custom_entity_parser_if_needed(dataset)
         language = dataset[LANGUAGE]
-        random_state = check_random_state(self.config.random_seed)
 
         data_augmentation_config = self.config.data_augmentation_config
         utterances, classes, intent_list = build_training_data(
             dataset, language, data_augmentation_config, self.resources,
-            random_state)
+            self.random_state)
 
         self.intent_list = intent_list
         if len(self.intent_list) <= 1:
@@ -81,7 +88,8 @@ def fit(self, dataset):
             config=self.config.featurizer_config,
             builtin_entity_parser=self.builtin_entity_parser,
             custom_entity_parser=self.custom_entity_parser,
-            resources=self.resources
+            resources=self.resources,
+            random_state=self.random_state,
         )
         self.featurizer.language = language
 
@@ -94,8 +102,8 @@ def fit(self, dataset):
             return self
 
         alpha = get_regularization_factor(dataset)
-        self.classifier = SGDClassifier(random_state=random_state,
-                                        alpha=alpha, **LOG_REG_ARGS)
+        self.classifier = SGDClassifier(
+            random_state=self.random_state, alpha=alpha, **LOG_REG_ARGS)
         self.classifier.fit(x, classes)
         logger.debug("%s", DifferedLoggingMessage(self.log_best_features))
         return self

snips_nlu/intent_parser/probabilistic_intent_parser.py

@@ -69,7 +69,9 @@ def fit(self, dataset, force_retrain=True):
                 self.config.intent_classifier_config,
                 builtin_entity_parser=self.builtin_entity_parser,
                 custom_entity_parser=self.custom_entity_parser,
-                resources=self.resources)
+                resources=self.resources,
+                random_state=self.random_state,
+            )
 
         if force_retrain or not self.intent_classifier.fitted:
             self.intent_classifier.fit(dataset)
@@ -85,7 +87,9 @@ def fit(self, dataset, force_retrain=True):
                     slot_filler_config,
                     builtin_entity_parser=self.builtin_entity_parser,
                     custom_entity_parser=self.custom_entity_parser,
-                    resources=self.resources)
+                    resources=self.resources,
+                    random_state=self.random_state,
+                )
             if force_retrain or not self.slot_fillers[intent_name].fitted:
                 self.slot_fillers[intent_name].fit(dataset, intent_name)
         logger.debug("Fitted slot fillers in %s",

snips_nlu/nlu_engine/nlu_engine.py

@@ -23,8 +23,9 @@
 from snips_nlu.entity_parser import CustomEntityParser
 from snips_nlu.entity_parser.builtin_entity_parser import (
     BuiltinEntityParser, is_builtin_entity)
-from snips_nlu.exceptions import InvalidInputError, IntentNotFoundError, \
-    LoadingError, IncompatibleModelError
+from snips_nlu.exceptions import (
+    InvalidInputError, IntentNotFoundError, LoadingError,
+    IncompatibleModelError)
 from snips_nlu.intent_parser import IntentParser
 from snips_nlu.pipeline.configs import NLUEngineConfig
 from snips_nlu.pipeline.processing_unit import ProcessingUnit
@@ -117,7 +118,9 @@ def fit(self, dataset, force_retrain=True):
                     parser_config,
                     builtin_entity_parser=self.builtin_entity_parser,
                     custom_entity_parser=self.custom_entity_parser,
-                    resources=self.resources)
+                    resources=self.resources,
+                    random_state=self.random_state,
+                )
 
             if force_retrain or not recycled_parser.fitted:
                 recycled_parser.fit(dataset, force_retrain)

snips_nlu/pipeline/configs/intent_classifier.py

@@ -13,16 +13,13 @@ class LogRegIntentClassifierConfig(FromDict, ProcessingUnitConfig):
     """Configuration of a :class:`.LogRegIntentClassifier`"""
 
     # pylint: disable=line-too-long
-    def __init__(self, data_augmentation_config=None, featurizer_config=None,
-                 random_seed=None):
+    def __init__(self, data_augmentation_config=None, featurizer_config=None):
         """
         Args:
             data_augmentation_config (:class:`IntentClassifierDataAugmentationConfig`):
                     Defines the strategy of the underlying data augmentation
             featurizer_config (:class:`FeaturizerConfig`): Configuration of the
                 :class:`.Featurizer` used underneath
-            random_seed (int, optional): Allows to fix the seed ot have
-                reproducible trainings
         """
         if data_augmentation_config is None:
             data_augmentation_config = IntentClassifierDataAugmentationConfig()
@@ -32,7 +29,6 @@ def __init__(self, data_augmentation_config=None, featurizer_config=None,
         self.data_augmentation_config = data_augmentation_config
         self._featurizer_config = None
         self.featurizer_config = featurizer_config
-        self.random_seed = random_seed
 
     # pylint: enable=line-too-long
 
@@ -83,8 +79,7 @@ def to_dict(self):
             "unit_name": self.unit_name,
             "data_augmentation_config":
                 self.data_augmentation_config.to_dict(),
-            "featurizer_config": self.featurizer_config.to_dict(),
-            "random_seed": self.random_seed
+            "featurizer_config": self.featurizer_config.to_dict()
         }