-
Notifications
You must be signed in to change notification settings - Fork 4.6k
/
ensemble.py
439 lines (362 loc) 路 16.6 KB
/
ensemble.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
import importlib
import json
import logging
import os
import sys
from collections import defaultdict
from datetime import datetime
from typing import Text, Optional, Any, List, Dict, Tuple
import numpy as np
import rasa.core
import rasa.utils.io
from rasa.constants import MINIMUM_COMPATIBLE_VERSION, DOCS_BASE_URL
from rasa.core import utils, training
from rasa.core.constants import USER_INTENT_BACK, USER_INTENT_RESTART
from rasa.core.actions.action import (
ACTION_LISTEN_NAME,
ACTION_BACK_NAME,
ACTION_RESTART_NAME,
)
from rasa.core.domain import Domain
from rasa.core.events import SlotSet, ActionExecuted, ActionExecutionRejected
from rasa.core.exceptions import UnsupportedDialogueModelError
from rasa.core.featurizers import MaxHistoryTrackerFeaturizer
from rasa.core.policies.policy import Policy
from rasa.core.policies.mapping_policy import MappingPolicy
from rasa.core.policies.fallback import FallbackPolicy
from rasa.core.policies.memoization import MemoizationPolicy, AugmentedMemoizationPolicy
from rasa.core.trackers import DialogueStateTracker
from rasa.core import registry
from rasa.utils.common import class_from_module_path
logger = logging.getLogger(__name__)
class PolicyEnsemble(object):
versioned_packages = ["rasa", "tensorflow", "sklearn"]
def __init__(
self, policies: List[Policy], action_fingerprints: Optional[Dict] = None
) -> None:
self.policies = policies
self.training_trackers = None
self.date_trained = None
if action_fingerprints:
self.action_fingerprints = action_fingerprints
else:
self.action_fingerprints = {}
self._check_priorities()
self._check_for_important_policies()
def _check_for_important_policies(self):
if not any([isinstance(policy, MappingPolicy) for policy in self.policies]):
logger.info(
"MappingPolicy not included in policy ensemble. Default intents "
"'{} and {} will not trigger actions '{}' and '{}'."
"".format(
USER_INTENT_RESTART,
USER_INTENT_BACK,
ACTION_RESTART_NAME,
ACTION_BACK_NAME,
)
)
@staticmethod
def _training_events_from_trackers(training_trackers):
events_metadata = defaultdict(set)
for t in training_trackers:
tracker = t.init_copy()
for event in t.events:
tracker.update(event)
if not isinstance(event, ActionExecuted):
action_name = tracker.latest_action_name
events_metadata[action_name].add(event)
return events_metadata
@staticmethod
def check_domain_ensemble_compatibility(
ensemble: Optional["PolicyEnsemble"], domain: Optional[Domain]
) -> None:
"""Check for elements that only work with certain policy/domain combinations."""
from rasa.core.policies.form_policy import FormPolicy
from rasa.core.policies.mapping_policy import MappingPolicy
from rasa.core.policies.two_stage_fallback import TwoStageFallbackPolicy
policies_needing_validation = [
FormPolicy,
MappingPolicy,
TwoStageFallbackPolicy,
]
for policy in policies_needing_validation:
policy.validate_against_domain(ensemble, domain)
def _check_priorities(self) -> None:
"""Checks for duplicate policy priorities within PolicyEnsemble."""
priority_dict = defaultdict(list)
for p in self.policies:
priority_dict[p.priority].append(type(p).__name__)
for k, v in priority_dict.items():
if len(v) > 1:
logger.warning(
(
"Found policies {} with same priority {} "
"in PolicyEnsemble. When personalizing "
"priorities, be sure to give all policies "
"different priorities. More information: "
"{}/core/policies/"
).format(v, k, DOCS_BASE_URL)
)
def train(
self,
training_trackers: List[DialogueStateTracker],
domain: Domain,
**kwargs: Any
) -> None:
if training_trackers:
for policy in self.policies:
policy.train(training_trackers, domain, **kwargs)
else:
logger.info("Skipped training, because there are no training samples.")
self.training_trackers = training_trackers
self.date_trained = datetime.now().strftime("%Y%m%d-%H%M%S")
def probabilities_using_best_policy(
self, tracker: DialogueStateTracker, domain: Domain
) -> Tuple[Optional[List[float]], Optional[Text]]:
raise NotImplementedError
def _max_histories(self) -> List[Optional[int]]:
"""Return max history."""
max_histories = []
for p in self.policies:
if isinstance(p.featurizer, MaxHistoryTrackerFeaturizer):
max_histories.append(p.featurizer.max_history)
else:
max_histories.append(None)
return max_histories
@staticmethod
def _create_action_fingerprints(training_events):
"""Fingerprint each action using the events it created during train.
This allows us to emit warnings when the model is used
if an action does things it hasn't done during training."""
if not training_events:
return None
action_fingerprints = {}
for k, vs in training_events.items():
slots = list({v.key for v in vs if isinstance(v, SlotSet)})
action_fingerprints[k] = {"slots": slots}
return action_fingerprints
def _add_package_version_info(self, metadata: Dict[Text, Any]) -> None:
"""Adds version info for self.versioned_packages to metadata."""
for package_name in self.versioned_packages:
try:
p = importlib.import_module(package_name)
v = p.__version__ # pytype: disable=attribute-error
metadata[package_name] = v
except ImportError:
pass
def _persist_metadata(
self, path: Text, dump_flattened_stories: bool = False
) -> None:
"""Persists the domain specification to storage."""
# make sure the directory we persist exists
domain_spec_path = os.path.join(path, "metadata.json")
training_data_path = os.path.join(path, "stories.md")
rasa.utils.io.create_directory_for_file(domain_spec_path)
policy_names = [utils.module_path_from_instance(p) for p in self.policies]
training_events = self._training_events_from_trackers(self.training_trackers)
action_fingerprints = self._create_action_fingerprints(training_events)
metadata = {
"action_fingerprints": action_fingerprints,
"python": ".".join([str(s) for s in sys.version_info[:3]]),
"max_histories": self._max_histories(),
"ensemble_name": self.__module__ + "." + self.__class__.__name__,
"policy_names": policy_names,
"trained_at": self.date_trained,
}
self._add_package_version_info(metadata)
utils.dump_obj_as_json_to_file(domain_spec_path, metadata)
# if there are lots of stories, saving flattened stories takes a long
# time, so this is turned off by default
if dump_flattened_stories:
training.persist_data(self.training_trackers, training_data_path)
def persist(self, path: Text, dump_flattened_stories: bool = False) -> None:
"""Persists the policy to storage."""
self._persist_metadata(path, dump_flattened_stories)
for i, policy in enumerate(self.policies):
dir_name = "policy_{}_{}".format(i, type(policy).__name__)
policy_path = os.path.join(path, dir_name)
policy.persist(policy_path)
@classmethod
def load_metadata(cls, path):
metadata_path = os.path.join(path, "metadata.json")
metadata = json.loads(rasa.utils.io.read_file(os.path.abspath(metadata_path)))
return metadata
@staticmethod
def ensure_model_compatibility(metadata, version_to_check=None):
from packaging import version
if version_to_check is None:
version_to_check = MINIMUM_COMPATIBLE_VERSION
model_version = metadata.get("rasa", "0.0.0")
if version.parse(model_version) < version.parse(version_to_check):
raise UnsupportedDialogueModelError(
"The model version is too old to be "
"loaded by this Rasa Core instance. "
"Either retrain the model, or run with "
"an older version. "
"Model version: {} Instance version: {} "
"Minimal compatible version: {}"
"".format(model_version, rasa.__version__, version_to_check),
model_version,
)
@classmethod
def _ensure_loaded_policy(cls, policy, policy_cls, policy_name: Text):
if policy is None:
raise Exception(
"Failed to load policy {}: load returned None".format(policy_name)
)
elif not isinstance(policy, policy_cls):
raise Exception(
"Failed to load policy {}: "
"load returned object that is not instance of its own class"
"".format(policy_name)
)
@classmethod
def load(cls, path: Text) -> "PolicyEnsemble":
"""Loads policy and domain specification from storage"""
metadata = cls.load_metadata(path)
cls.ensure_model_compatibility(metadata)
policies = []
for i, policy_name in enumerate(metadata["policy_names"]):
policy_cls = registry.policy_from_module_path(policy_name)
dir_name = "policy_{}_{}".format(i, policy_cls.__name__)
policy_path = os.path.join(path, dir_name)
policy = policy_cls.load(policy_path)
cls._ensure_loaded_policy(policy, policy_cls, policy_name)
policies.append(policy)
ensemble_cls = class_from_module_path(metadata["ensemble_name"])
fingerprints = metadata.get("action_fingerprints", {})
ensemble = ensemble_cls(policies, fingerprints)
return ensemble
@classmethod
def from_dict(cls, dictionary: Dict[Text, Any]) -> List[Policy]:
policies = dictionary.get("policies") or dictionary.get("policy")
if policies is None:
raise InvalidPolicyConfig(
"You didn't define any policies. "
"Please define them under 'policies:' "
"in your policy configuration file."
)
if len(policies) == 0:
raise InvalidPolicyConfig(
"The policy configuration file has to include at least one policy."
)
parsed_policies = []
for policy in policies:
policy_name = policy.pop("name")
if policy.get("featurizer"):
featurizer_func, featurizer_config = cls.get_featurizer_from_dict(
policy
)
if featurizer_config.get("state_featurizer"):
state_featurizer_func, state_featurizer_config = cls.get_state_featurizer_from_dict(
featurizer_config
)
# override featurizer's state_featurizer
# with real state_featurizer class
featurizer_config["state_featurizer"] = state_featurizer_func(
**state_featurizer_config
)
# override policy's featurizer with real featurizer class
policy["featurizer"] = featurizer_func(**featurizer_config)
try:
constr_func = registry.policy_from_module_path(policy_name)
policy_object = constr_func(**policy)
parsed_policies.append(policy_object)
except (ImportError, AttributeError):
raise InvalidPolicyConfig(
"Module for policy '{}' could not "
"be loaded. Please make sure the "
"name is a valid policy."
"".format(policy_name)
)
return parsed_policies
@classmethod
def get_featurizer_from_dict(cls, policy):
# policy can have only 1 featurizer
if len(policy["featurizer"]) > 1:
raise InvalidPolicyConfig("policy can have only 1 featurizer")
featurizer_config = policy["featurizer"][0]
featurizer_name = featurizer_config.pop("name")
featurizer_func = registry.featurizer_from_module_path(featurizer_name)
return featurizer_func, featurizer_config
@classmethod
def get_state_featurizer_from_dict(cls, featurizer_config):
# featurizer can have only 1 state featurizer
if len(featurizer_config["state_featurizer"]) > 1:
raise InvalidPolicyConfig("featurizer can have only 1 state featurizer")
state_featurizer_config = featurizer_config["state_featurizer"][0]
state_featurizer_name = state_featurizer_config.pop("name")
state_featurizer_func = registry.featurizer_from_module_path(
state_featurizer_name
)
return state_featurizer_func, state_featurizer_config
def continue_training(
self, trackers: List[DialogueStateTracker], domain: Domain, **kwargs: Any
) -> None:
self.training_trackers.extend(trackers)
for p in self.policies:
p.continue_training(self.training_trackers, domain, **kwargs)
class SimplePolicyEnsemble(PolicyEnsemble):
@staticmethod
def is_not_memo_policy(best_policy_name):
is_memo = best_policy_name.endswith("_" + MemoizationPolicy.__name__)
is_augmented = best_policy_name.endswith(
"_" + AugmentedMemoizationPolicy.__name__
)
return not (is_memo or is_augmented)
def probabilities_using_best_policy(
self, tracker: DialogueStateTracker, domain: Domain
) -> Tuple[Optional[List[float]], Optional[Text]]:
result = None
max_confidence = -1
best_policy_name = None
best_policy_priority = -1
for i, p in enumerate(self.policies):
probabilities = p.predict_action_probabilities(tracker, domain)
if len(tracker.events) > 0 and isinstance(
tracker.events[-1], ActionExecutionRejected
):
probabilities[
domain.index_for_action(tracker.events[-1].action_name)
] = 0.0
confidence = np.max(probabilities)
if (confidence, p.priority) > (max_confidence, best_policy_priority):
max_confidence = confidence
result = probabilities
best_policy_name = "policy_{}_{}".format(i, type(p).__name__)
best_policy_priority = p.priority
if (
result is not None
and result.index(max_confidence)
== domain.index_for_action(ACTION_LISTEN_NAME)
and tracker.latest_action_name == ACTION_LISTEN_NAME
and self.is_not_memo_policy(best_policy_name)
):
# Trigger the fallback policy when ActionListen is predicted after
# a user utterance. This is done on the condition that:
# - a fallback policy is present,
# - there was just a user message and the predicted
# action is action_listen by a policy
# other than the MemoizationPolicy
fallback_idx_policy = [
(i, p)
for i, p in enumerate(self.policies)
if isinstance(p, FallbackPolicy)
]
if fallback_idx_policy:
fallback_idx, fallback_policy = fallback_idx_policy[0]
logger.debug(
"Action 'action_listen' was predicted after "
"a user message using {}. "
"Predicting fallback action: {}"
"".format(best_policy_name, fallback_policy.fallback_action_name)
)
result = fallback_policy.fallback_scores(domain)
best_policy_name = "policy_{}_{}".format(
fallback_idx, type(fallback_policy).__name__
)
logger.debug("Predicted next action using {}".format(best_policy_name))
return result, best_policy_name
class InvalidPolicyConfig(Exception):
"""Exception that can be raised when policy config is not valid."""
pass