def test_normalization(
self,
trained_policy: Policy,
tracker: DialogueStateTracker,
default_domain: Domain,
monkeypatch: MonkeyPatch,
):
precomputations = None
# first check the output is what we expect
predicted_probabilities = trained_policy.predict_action_probabilities(
tracker,
default_domain,
precomputations,
).probabilities
output_sums_to_1 = sum(predicted_probabilities) == pytest.approx(1)
assert output_sums_to_1
# also check our function is not called
mock = Mock()
monkeypatch.setattr(train_utils, "normalize", mock.normalize)
trained_policy.predict_action_probabilities(
tracker,
default_domain,
precomputations,
)
mock.normalize.assert_not_called()
def test_normalization(
self,
trained_policy: Policy,
tracker: DialogueStateTracker,
default_domain: Domain,
monkeypatch: MonkeyPatch,
):
precomputations = None
# first check the output is what we expect
predicted_probabilities = trained_policy.predict_action_probabilities(
tracker,
default_domain,
precomputations,
).probabilities
# there should be no normalization
assert all([confidence > 0 for confidence in predicted_probabilities])
# also check our function is not called
mock = Mock()
monkeypatch.setattr(train_utils, "normalize", mock.normalize)
trained_policy.predict_action_probabilities(
tracker,
default_domain,
precomputations,
)
mock.normalize.assert_not_called()
def test_persist_and_load(
self,
trained_policy: PolicyGraphComponent,
default_domain: Domain,
should_finetune: bool,
stories_path: Text,
model_storage: ModelStorage,
resource: Resource,
execution_context: ExecutionContext,
):
loaded = trained_policy.__class__.load(
self._config(trained_policy.config),
model_storage,
resource,
dataclasses.replace(execution_context,
is_finetuning=should_finetune),
)
assert loaded.finetune_mode == should_finetune
trackers = train_trackers(default_domain,
stories_path,
augmentation_factor=20)
for tracker in trackers:
predicted_probabilities = loaded.predict_action_probabilities(
tracker, default_domain)
actual_probabilities = trained_policy.predict_action_probabilities(
tracker, default_domain)
assert predicted_probabilities == actual_probabilities
def _get_next_action(
policy: PolicyGraphComponent, events: List[Event], domain: Domain
) -> Text:
tracker = get_tracker(events)
scores = policy.predict_action_probabilities(tracker, domain,).probabilities
index = scores.index(max(scores))
return domain.action_names_or_texts[index]
def test_normalization(
self,
trained_policy: Policy,
tracker: DialogueStateTracker,
default_domain: Domain,
monkeypatch: MonkeyPatch,
):
# Mock actual normalization method
mock = Mock()
monkeypatch.setattr(train_utils, "normalize", mock.normalize)
trained_policy.predict_action_probabilities(
tracker,
default_domain,
precomputations=None,
)
# function should not get called for margin loss_type
mock.normalize.assert_not_called()
def test_prediction_on_empty_tracker(self, trained_policy: Policy,
default_domain: Domain):
tracker = DialogueStateTracker(DEFAULT_SENDER_ID, default_domain.slots)
prediction = trained_policy.predict_action_probabilities(
tracker,
default_domain,
precomputations=None,
)
assert not prediction.is_end_to_end_prediction
assert len(prediction.probabilities) == default_domain.num_actions
assert max(prediction.probabilities) <= 1.0
assert min(prediction.probabilities) >= 0.0