def test_fallback_mapping_restart():
domain = Domain.load("data/test_domains/default.yml")
events = [
ActionExecuted(ACTION_DEFAULT_FALLBACK_NAME),
utilities.user_uttered(USER_INTENT_RESTART, 1),
]
tracker = DialogueStateTracker.from_events("test", events, [])
two_stage_fallback_policy = TwoStageFallbackPolicy(
priority=2, deny_suggestion_intent_name="deny"
)
mapping_policy = MappingPolicy(priority=1)
mapping_fallback_ensemble = SimplePolicyEnsemble(
[two_stage_fallback_policy, mapping_policy]
)
result, best_policy = mapping_fallback_ensemble.probabilities_using_best_policy(
tracker, domain, RegexInterpreter()
)
max_confidence_index = result.index(max(result))
index_of_mapping_policy = 1
next_action = domain.action_for_index(max_confidence_index, None)
assert best_policy == f"policy_{index_of_mapping_policy}_{MappingPolicy.__name__}"
assert next_action.name() == ACTION_RESTART_NAME
def test_mapping_wins_over_form(events: List[Event]):
domain = """
forms:
- test-form
"""
domain = Domain.from_yaml(domain)
tracker = DialogueStateTracker.from_events("test", events, [])
ensemble = SimplePolicyEnsemble(
[
MappingPolicy(),
ConstantPolicy(priority=1, predict_index=0),
FormPolicy(),
FallbackPolicy(),
]
)
result, best_policy = ensemble.probabilities_using_best_policy(
tracker, domain, RegexInterpreter()
)
max_confidence_index = result.index(max(result))
next_action = domain.action_for_index(max_confidence_index, None)
index_of_mapping_policy = 0
assert best_policy == f"policy_{index_of_mapping_policy}_{MappingPolicy.__name__}"
assert next_action.name() == ACTION_RESTART_NAME
def train_core(
domain_file: Text = "domain.yml",
model_directory: Text = "models",
model_name: Text = "current",
training_data_file: Text = "data/stories.md",
):
agent = Agent(
domain_file,
policies=[
MemoizationPolicy(max_history=3),
MappingPolicy(),
RestaurantPolicy(batch_size=100, epochs=100, validation_split=0.2),
],
)
# augmentation_factor 扩展系数 10
training_data = agent.load_data(training_data_file, augmentation_factor=10)
agent.train(training_data)
# Attention: agent.persist stores the model and all meta data into a folder.
# The folder itself is not zipped. 未打包的的模型文件
model_path = os.path.join(model_directory, model_name, "core")
agent.persist(model_path)
logger.info("Model trained. Stored in '{}'.".format(model_path))
return model_path
async def train_core(domain_file="domain.yml",
model_directory="models",
model_name="current",
training_data_file="data/stories.md"):
agent = Agent(
domain_file,
policies=[
MemoizationPolicy(max_history=3),
MappingPolicy(),
RestaurantPolicy(batch_size=100, epochs=100, validation_split=0.2),
],
)
training_data_file = "data/tiny_stories.md"
training_data = await agent.load_data(training_data_file,
augmentation_factor=10)
# show_training_data(training_data)
# print(type(training_data))
# print(training_data)
for data in training_data:
print(type(data))
# viz_domain(data.domain)
# viz_TrackerWithCachedStates(data, view_domain=False)
# exit()
agent.train(training_data)
# Attention: agent.persist stores the model and all meta data into a folder.
# The folder itself is not zipped.
model_path = os.path.join(model_directory, model_name, "core")
agent.persist(model_path)
logger.info(f"Model trained. Stored in '{model_path}'.")
return model_path
async def test_maping_policy():
default_domain = Domain.load("{}/data/default_with_mapping.yml".format(prj_dir))
policy = MappingPolicy()
events = [
ActionExecuted(ACTION_LISTEN_NAME),
user_uttered(intent_mapping[0][0], 1),
]
tracker = DialogueStateTracker.from_events("sender", events, [], 20)
viz_tracker(tracker)
scores = policy.predict_action_probabilities(tracker, default_domain)
print(default_domain.action_names)
class_to_idx = {cl:idx for idx, cl in enumerate(default_domain.action_names)}
print(class_to_idx)
index = scores.index(max(scores))
print("action names: {}".format(default_domain.action_names))
print(default_domain.action_names[index])
async def train_core(domain_file, training_data_file, model_directory):
agent = Agent(domain_file, policies=[MemoizationPolicy(max_history=3), MappingPolicy(), KerasPolicy(epochs=500)])
training_data = await agent.load_data(training_data_file, augmentation_factor=10)
agent.train(training_data)
# Attention: agent.persist stores the model and all meta data into a folder.
# The folder itself is not zipped.
model_path = os.path.join(model_directory, "core")
agent.persist(model_path)
logger.info(f"Model trained. Stored in '{model_path}'.")
return model_path
async def train_dialogue(domain_file="domain.yml",
model_path="models/dialogue",
training_data_file="data/stories.md"):
agent = Agent(domain_file,
policies=[MemoizationPolicy(max_history=3),
MappingPolicy(),
RestaurantPolicy(batch_size=100, epochs=400,
validation_split=0.2)])
training_data = await agent.load_data(training_data_file)
agent.train(
training_data
)
agent.persist(model_path)
return agent
def test_fallback_wins_over_mapping():
domain = Domain.load("data/test_domains/default.yml")
events = [
ActionExecuted(ACTION_LISTEN_NAME),
# Low confidence should trigger fallback
utilities.user_uttered(USER_INTENT_RESTART, 0.0001),
]
tracker = DialogueStateTracker.from_events("test", events, [])
ensemble = SimplePolicyEnsemble([FallbackPolicy(), MappingPolicy()])
result, best_policy = ensemble.probabilities_using_best_policy(
tracker, domain)
max_confidence_index = result.index(max(result))
index_of_fallback_policy = 0
next_action = domain.action_for_index(max_confidence_index, None)
assert best_policy == f"policy_{index_of_fallback_policy}_{FallbackPolicy.__name__}"
assert next_action.name() == ACTION_DEFAULT_FALLBACK_NAME
async def train_core(
domain_file: Text = "domain.yml",
model_path: Text = "models/core",
training_data_file: Text = "data/stories.md",
):
agent = Agent(
domain_file,
policies=[
MemoizationPolicy(max_history=3),
MappingPolicy(),
RestaurantPolicy(batch_size=100, epochs=400, validation_split=0.2),
],
)
training_data = await agent.load_data(training_data_file)
agent.train(training_data)
# Attention: agent.persist stores the model and all meta data into a folder.
# The folder itself is not zipped.
agent.persist(model_path)
logger.info("Model trained. Stored in '{}'.".format(model_path))
return model_path
index_of_mapping_policy = 0
assert best_policy == f"policy_{index_of_mapping_policy}_{MappingPolicy.__name__}"
assert next_action.name() == ACTION_RESTART_NAME
@pytest.mark.parametrize(
"ensemble",
[
SimplePolicyEnsemble(
[
FormPolicy(),
ConstantPolicy(FORM_POLICY_PRIORITY - 1, 0),
FallbackPolicy(),
]
),
SimplePolicyEnsemble([FormPolicy(), MappingPolicy()]),
],
)
def test_form_wins_over_everything_else(ensemble: SimplePolicyEnsemble):
form_name = "test-form"
domain = f"""
forms:
- {form_name}
"""
domain = Domain.from_yaml(domain)
events = [
Form("test-form"),
ActionExecuted(ACTION_LISTEN_NAME),
utilities.user_uttered("test", 1),
]
def create_policy(self, featurizer, priority):
p = MappingPolicy()
return p
def create_policy(
self, featurizer: Optional[TrackerFeaturizer], priority: int
) -> Policy:
return MappingPolicy()
