def training_states_and_actions(
self, trackers: List[DialogueStateTracker], domain: Domain
) -> Tuple[List[List[Optional[Dict[Text, float]]]], List[List[Text]]]:
"""Transforms list of trackers to lists of states and actions.
Training data is padded up to the max_history with -1.
"""
trackers_as_states = []
trackers_as_actions = []
# from multiple states that create equal featurizations
# we only need to keep one.
hashed_examples = set()
logger.debug("Creating states and action examples from "
"collected trackers (by {}({}))..."
"".format(
type(self).__name__,
type(self.state_featurizer).__name__))
pbar = tqdm(trackers,
desc="Processed trackers",
disable=is_logging_disabled())
for tracker in pbar:
states = self._create_states(tracker, domain, True)
idx = 0
for event in tracker.applied_events():
if isinstance(event, ActionExecuted):
if not event.unpredictable:
# only actions which can be
# predicted at a stories start
sliced_states = self.slice_state_history(
states[:idx + 1], self.max_history)
if self.remove_duplicates:
hashed = self._hash_example(
sliced_states, event.action_name)
# only continue with tracker_states that created a
# hashed_featurization we haven't observed
if hashed not in hashed_examples:
hashed_examples.add(hashed)
trackers_as_states.append(sliced_states)
trackers_as_actions.append([event.action_name])
else:
trackers_as_states.append(sliced_states)
trackers_as_actions.append([event.action_name])
pbar.set_postfix({
"# actions":
"{:d}".format(len(trackers_as_actions))
})
idx += 1
logger.debug("Created {} action examples.".format(
len(trackers_as_actions)))
return trackers_as_states, trackers_as_actions
def training_states_and_actions(
self, trackers: List[DialogueStateTracker], domain: Domain
) -> Tuple[List[List[State]], List[List[Text]]]:
"""Transforms list of trackers to lists of states and actions.
Training data is padded up to the length of the longest dialogue with -1.
Args:
trackers: The trackers to transform
domain: The domain
Returns:
A tuple of list of states and list of actions.
"""
trackers_as_states = []
trackers_as_actions = []
logger.debug(
"Creating states and action examples from "
"collected trackers (by {}({}))..."
"".format(type(self).__name__, type(self.state_featurizer).__name__)
)
pbar = tqdm(
trackers,
desc="Processed trackers",
disable=common_utils.is_logging_disabled(),
)
for tracker in pbar:
states = self._create_states(tracker, domain)
delete_first_state = False
actions = []
for event in tracker.applied_events():
if not isinstance(event, ActionExecuted):
continue
if not event.unpredictable:
# only actions which can be
# predicted at a stories start
actions.append(event.action_name or event.action_text)
else:
# unpredictable actions can be
# only the first in the story
if delete_first_state:
raise InvalidStory(
f"Found two unpredictable actions in one story "
f"'{tracker.sender_id}'. Check your story files."
)
delete_first_state = True
if delete_first_state:
states = states[1:]
trackers_as_states.append(states[:-1])
trackers_as_actions.append(actions)
return trackers_as_states, trackers_as_actions
def training_states_and_actions(
self, trackers: List[DialogueStateTracker],
domain: Domain) -> Tuple[List[List[Dict]], List[List[Text]]]:
"""Transforms list of trackers to lists of states and actions.
Training data is padded up to the length of the longest dialogue with -1.
"""
trackers_as_states = []
trackers_as_actions = []
logger.debug("Creating states and action examples from "
"collected trackers (by {}({}))..."
"".format(
type(self).__name__,
type(self.state_featurizer).__name__))
pbar = tqdm(trackers,
desc="Processed trackers",
disable=is_logging_disabled())
for tracker in pbar:
states = self._create_states(tracker,
domain,
is_binary_training=True)
delete_first_state = False
actions = []
for event in tracker.applied_events():
if isinstance(event, ActionExecuted):
if not event.unpredictable:
# only actions which can be
# predicted at a stories start
actions.append(event.action_name)
else:
# unpredictable actions can be
# only the first in the story
if delete_first_state:
raise Exception("Found two unpredictable "
"actions in one story."
"Check your story files.")
else:
delete_first_state = True
if delete_first_state:
states = states[1:]
trackers_as_states.append(states[:-1])
trackers_as_actions.append(actions)
self.max_len = self._calculate_max_len(trackers_as_actions)
logger.debug("The longest dialogue has {} actions.".format(
self.max_len))
return trackers_as_states, trackers_as_actions
def _create_lookup_from_states(
self,
trackers_as_states: List[List[Dict]],
trackers_as_actions: List[List[Text]],
) -> Dict[Text, Text]:
"""Creates lookup dictionary from the tracker represented as states.
Args:
trackers_as_states: representation of the trackers as a list of states
trackers_as_actions: representation of the trackers as a list of actions
Returns:
lookup dictionary
"""
lookup = {}
if not trackers_as_states:
return lookup
if self.max_history:
assert len(trackers_as_states[0]) == self.max_history, (
f"Trying to memorizefeaturized data with {len(trackers_as_states[0])} "
f"historic turns. Expected: {self.max_history}")
assert len(trackers_as_actions[0]) == 1, (
f"The second dimension of trackers_as_action should be 1, "
f"instead of {len(trackers_as_actions[0])}")
ambiguous_feature_keys = set()
pbar = tqdm(
zip(trackers_as_states, trackers_as_actions),
desc="Processed actions",
disable=is_logging_disabled(),
)
for states, actions in pbar:
action = actions[0]
feature_key = self._create_feature_key(states)
if feature_key not in ambiguous_feature_keys:
if feature_key in lookup.keys():
if lookup[feature_key] != action:
# delete contradicting example created by
# partial history augmentation from memory
ambiguous_feature_keys.add(feature_key)
del lookup[feature_key]
else:
lookup[feature_key] = action
pbar.set_postfix({"# examples": "{:d}".format(len(lookup))})
return lookup
def _add_states_to_lookup(
self, trackers_as_states, trackers_as_actions, domain, online=False
) -> None:
"""Add states to lookup dict"""
if not trackers_as_states:
return
assert len(trackers_as_states[0]) == self.max_history, (
"Trying to mem featurized data with {} historic turns. Expected: "
"{}".format(len(trackers_as_states[0]), self.max_history)
)
assert len(trackers_as_actions[0]) == 1, (
"The second dimension of trackers_as_action should be 1, "
"instead of {}".format(len(trackers_as_actions[0]))
)
ambiguous_feature_keys = set()
pbar = tqdm(
zip(trackers_as_states, trackers_as_actions),
desc="Processed actions",
disable=is_logging_disabled(),
)
for states, actions in pbar:
action = actions[0]
feature_key = self._create_feature_key(states)
feature_item = domain.index_for_action(action)
if feature_key not in ambiguous_feature_keys:
if feature_key in self.lookup.keys():
if self.lookup[feature_key] != feature_item:
if online:
logger.info(
"Original stories are "
"different for {} -- {}\n"
"Memorized the new ones for "
"now. Delete contradicting "
"examples after exporting "
"the new stories."
"".format(states, action)
)
self.lookup[feature_key] = feature_item
else:
# delete contradicting example created by
# partial history augmentation from memory
ambiguous_feature_keys.add(feature_key)
del self.lookup[feature_key]
else:
self.lookup[feature_key] = feature_item
pbar.set_postfix({"# examples": "{:d}".format(len(self.lookup))})
def generate(self) -> List[TrackerWithCachedStates]:
if self.config.remove_duplicates and self.config.unique_last_num_states:
logger.debug("Generated trackers will be deduplicated "
"based on their unique last {} states."
"".format(self.config.unique_last_num_states))
self._mark_first_action_in_story_steps_as_unpredictable()
active_trackers = defaultdict(list)
init_tracker = TrackerWithCachedStates(
"",
self.domain.slots,
max_event_history=self.config.tracker_limit,
domain=self.domain,
)
active_trackers[STORY_START].append(init_tracker)
# trackers that are sent to a featurizer
finished_trackers = []
# keep story end trackers separately for augmentation
story_end_trackers = []
phase = 0 # one phase is one traversal of all story steps.
min_num_aug_phases = 3 if self.config.augmentation_factor > 0 else 0
logger.debug(
"Number of augmentation rounds is {}".format(min_num_aug_phases))
# placeholder to track gluing process of checkpoints
used_checkpoints = set()
previous_unused = set()
everything_reachable_is_reached = False
# we will continue generating data until we have reached all
# checkpoints that seem to be reachable. This is a heuristic,
# if we did not reach any new checkpoints in an iteration, we
# assume we have reached all and stop.
while not everything_reachable_is_reached or phase < min_num_aug_phases:
phase_name = self._phase_name(everything_reachable_is_reached,
phase)
num_active_trackers = self._count_trackers(active_trackers)
if num_active_trackers:
logger.debug("Starting {} ... (with {} trackers)"
"".format(phase_name, num_active_trackers))
else:
logger.debug("There are no trackers for {}".format(phase_name))
break
# track unused checkpoints for this phase
unused_checkpoints = set() # type: Set[Text]
pbar = tqdm(
self.story_graph.ordered_steps(),
desc="Processed Story Blocks",
disable=is_logging_disabled(),
)
for step in pbar:
incoming_trackers = [] # type: List[TrackerWithCachedStates]
for start in step.start_checkpoints:
if active_trackers[start.name]:
ts = start.filter_trackers(active_trackers[start.name])
incoming_trackers.extend(ts)
used_checkpoints.add(start.name)
elif start.name not in used_checkpoints:
# need to skip - there was no previous step that
# had this start checkpoint as an end checkpoint
# it will be processed in next phases
unused_checkpoints.add(start.name)
if not incoming_trackers:
# if there are no trackers,
# we can skip the rest of the loop
continue
# these are the trackers that reached this story
# step and that need to handle all events of the step
if self.config.remove_duplicates:
incoming_trackers, end_trackers = self._remove_duplicate_trackers(
incoming_trackers)
# append end trackers to finished trackers
finished_trackers.extend(end_trackers)
if everything_reachable_is_reached:
# augmentation round
incoming_trackers = self._subsample_trackers(
incoming_trackers,
self.config.max_number_of_augmented_trackers)
# update progress bar
pbar.set_postfix(
{"# trackers": "{:d}".format(len(incoming_trackers))})
trackers, end_trackers = self._process_step(
step, incoming_trackers)
# add end trackers to finished trackers
finished_trackers.extend(end_trackers)
# update our tracker dictionary with the trackers
# that handled the events of the step and
# that can now be used for further story steps
# that start with the checkpoint this step ended with
for end in step.end_checkpoints:
start_name = self._find_start_checkpoint_name(end.name)
active_trackers[start_name].extend(trackers)
if start_name in used_checkpoints:
# add end checkpoint as unused
# if this checkpoint was processed as
# start one before
unused_checkpoints.add(start_name)
if not step.end_checkpoints:
unique_ends = self._remove_duplicate_story_end_trackers(
trackers)
story_end_trackers.extend(unique_ends)
num_finished = len(finished_trackers) + len(story_end_trackers)
logger.debug("Finished phase ({} training samples found).".format(
num_finished))
# prepare next round
phase += 1
if not everything_reachable_is_reached:
# check if we reached all nodes that can be reached
# if we reached at least one more node this round
# than last one, we assume there is still
# something left to reach and we continue
unused_checkpoints = self._add_unused_end_checkpoints(
set(active_trackers.keys()), unused_checkpoints,
used_checkpoints)
active_trackers = self._filter_active_trackers(
active_trackers, unused_checkpoints)
num_active_trackers = self._count_trackers(active_trackers)
everything_reachable_is_reached = (unused_checkpoints
== previous_unused
or num_active_trackers == 0)
previous_unused = unused_checkpoints
if everything_reachable_is_reached:
# should happen only once
previous_unused -= used_checkpoints
# add trackers with unused checkpoints
# to finished trackers
for start_name in previous_unused:
finished_trackers.extend(active_trackers[start_name])
logger.debug("Data generation rounds finished.")
logger.debug("Found {} unused checkpoints".format(
len(previous_unused)))
phase = 0
else:
logger.debug("Found {} unused checkpoints "
"in current phase."
"".format(len(unused_checkpoints)))
logger.debug("Found {} active trackers "
"for these checkpoints."
"".format(num_active_trackers))
if everything_reachable_is_reached:
# augmentation round, so we process only
# story end checkpoints
# reset used checkpoints
used_checkpoints = set() # type: Set[Text]
# generate active trackers for augmentation
active_trackers = self._create_start_trackers_for_augmentation(
story_end_trackers)
finished_trackers.extend(story_end_trackers)
self._issue_unused_checkpoint_notification(previous_unused)
logger.debug("Found {} training trackers.".format(
len(finished_trackers)))
if self.config.augmentation_factor > 0:
augmented_trackers, original_trackers = [], []
for t in finished_trackers:
if t.is_augmented:
augmented_trackers.append(t)
else:
original_trackers.append(t)
augmented_trackers = self._subsample_trackers(
augmented_trackers,
self.config.max_number_of_augmented_trackers)
logger.debug("Subsampled to {} augmented training trackers."
"".format(len(augmented_trackers)))
logger.debug("There are {} original trackers.".format(
len(original_trackers)))
finished_trackers = original_trackers + augmented_trackers
return finished_trackers
def fit(
self,
model_data: RasaModelData,
epochs: int,
batch_size: Union[List[int], int],
evaluate_on_num_examples: int,
evaluate_every_num_epochs: int,
batch_strategy: Text,
silent: bool = False,
eager: bool = False,
) -> None:
"""Fit model data"""
tf.random.set_seed(self.random_seed)
np.random.seed(self.random_seed)
disable = silent or is_logging_disabled()
evaluation_model_data = None
if evaluate_on_num_examples > 0:
if not disable:
logger.info(
f"Validation accuracy is calculated every "
f"{evaluate_every_num_epochs} epochs."
)
model_data, evaluation_model_data = model_data.split(
evaluate_on_num_examples, self.random_seed
)
(
train_dataset_function,
tf_train_on_batch_function,
) = self._get_tf_train_functions(eager, model_data, batch_strategy)
(
evaluation_dataset_function,
tf_evaluation_on_batch_function,
) = self._get_tf_evaluation_functions(eager, evaluation_model_data)
val_results = {} # validation is not performed every epoch
progress_bar = tqdm(range(epochs), desc="Epochs", disable=disable)
training_steps = 0
for epoch in progress_bar:
epoch_batch_size = self.linearly_increasing_batch_size(
epoch, batch_size, epochs
)
training_steps = self._batch_loop(
train_dataset_function,
tf_train_on_batch_function,
epoch_batch_size,
True,
training_steps,
self.train_summary_writer,
)
if self.tensorboard_log_on_epochs:
self._log_metrics_for_tensorboard(epoch, self.train_summary_writer)
postfix_dict = self._get_metric_results()
if evaluate_on_num_examples > 0:
if self._should_evaluate(evaluate_every_num_epochs, epochs, epoch):
self._batch_loop(
evaluation_dataset_function,
tf_evaluation_on_batch_function,
epoch_batch_size,
False,
training_steps,
self.test_summary_writer,
)
if self.tensorboard_log_on_epochs:
self._log_metrics_for_tensorboard(
epoch, self.test_summary_writer
)
val_results = self._get_metric_results(prefix="val_")
postfix_dict.update(val_results)
progress_bar.set_postfix(postfix_dict)
if self.model_summary_file is not None:
self._write_model_summary()
self._training = None # training phase should be defined when building a graph
if not disable:
logger.info("Finished training.")
def _train_tf(
self,
X: np.ndarray,
Y: np.ndarray,
intents_for_X: np.ndarray,
loss: "Tensor",
is_training: "Tensor",
train_op: "Tensor",
) -> None:
"""Train tf graph"""
self.session.run(tf.global_variables_initializer())
if self.evaluate_on_num_examples:
logger.info("Accuracy is updated every {} epochs"
"".format(self.evaluate_every_num_epochs))
pbar = tqdm(range(self.epochs),
desc="Epochs",
disable=is_logging_disabled())
train_acc = 0
last_loss = 0
for ep in pbar:
indices = np.random.permutation(len(X))
batch_size = self._linearly_increasing_batch_size(ep)
batches_per_epoch = len(X) // batch_size + int(
len(X) % batch_size > 0)
ep_loss = 0
for i in range(batches_per_epoch):
end_idx = (i + 1) * batch_size
start_idx = i * batch_size
batch_a = X[indices[start_idx:end_idx]]
batch_pos_b = Y[indices[start_idx:end_idx]]
intents_for_b = intents_for_X[indices[start_idx:end_idx]]
# add negatives
batch_b = self._create_batch_b(batch_pos_b, intents_for_b)
sess_out = self.session.run(
{
"loss": loss,
"train_op": train_op
},
feed_dict={
self.a_in: batch_a,
self.b_in: batch_b,
is_training: True,
},
)
ep_loss += sess_out.get("loss") / batches_per_epoch
if self.evaluate_on_num_examples:
if (ep == 0 or (ep + 1) % self.evaluate_every_num_epochs == 0
or (ep + 1) == self.epochs):
train_acc = self._output_training_stat(
X, intents_for_X, is_training)
last_loss = ep_loss
pbar.set_postfix({
"loss": "{:.3f}".format(ep_loss),
"acc": "{:.3f}".format(train_acc),
})
else:
pbar.set_postfix({"loss": "{:.3f}".format(ep_loss)})
if self.evaluate_on_num_examples:
logger.info("Finished training embedding classifier, "
"loss={:.3f}, train accuracy={:.3f}"
"".format(last_loss, train_acc))
def train_tf_dataset(
train_init_op: "tf.Operation",
eval_init_op: "tf.Operation",
batch_size_in: "tf.Tensor",
loss: "tf.Tensor",
acc: "tf.Tensor",
train_op: "tf.Tensor",
session: "tf.Session",
is_training: "tf.Session",
epochs: int,
batch_size: Union[List[int], int],
evaluate_on_num_examples: int,
evaluate_every_num_epochs: int,
) -> None:
"""Train tf graph"""
session.run(tf.global_variables_initializer())
if evaluate_on_num_examples:
logger.info("Validation accuracy is calculated every {} epochs"
"".format(evaluate_every_num_epochs))
pbar = tqdm(range(epochs), desc="Epochs", disable=is_logging_disabled())
train_loss = 0
train_acc = 0
val_loss = 0
val_acc = 0
for ep in pbar:
ep_batch_size = linearly_increasing_batch_size(ep, batch_size, epochs)
session.run(train_init_op, feed_dict={batch_size_in: ep_batch_size})
ep_train_loss = 0
ep_train_acc = 0
batches_per_epoch = 0
while True:
try:
_, batch_train_loss, batch_train_acc = session.run(
[train_op, loss, acc], feed_dict={is_training: True})
batches_per_epoch += 1
ep_train_loss += batch_train_loss
ep_train_acc += batch_train_acc
except tf.errors.OutOfRangeError:
break
train_loss = ep_train_loss / batches_per_epoch
train_acc = ep_train_acc / batches_per_epoch
postfix_dict = {
"loss": "{:.3f}".format(train_loss),
"acc": "{:.3f}".format(train_acc),
}
if eval_init_op is not None:
if (ep + 1) % evaluate_every_num_epochs == 0 or (ep + 1) == epochs:
val_loss, val_acc = output_validation_stat(
eval_init_op,
loss,
acc,
session,
is_training,
batch_size_in,
ep_batch_size,
)
postfix_dict.update({
"val_loss": "{:.3f}".format(val_loss),
"val_acc": "{:.3f}".format(val_acc),
})
pbar.set_postfix(postfix_dict)
final_message = ("Finished training embedding policy, "
"train loss={:.3f}, train accuracy={:.3f}"
"".format(train_loss, train_acc))
if eval_init_op is not None:
final_message += (
", validation loss={:.3f}, validation accuracy={:.3f}"
"".format(val_loss, val_acc))
logger.info(final_message)
def training_states_and_actions(
self, trackers: List[DialogueStateTracker], domain: Domain
) -> Tuple[List[List[State]], List[List[Text]]]:
"""Transforms list of trackers to lists of states and actions.
Training data is padded up to the length of the longest dialogue with -1.
Args:
trackers: The trackers to transform
domain: The domain
Returns:
A tuple of list of states and list of actions.
"""
trackers_as_states = []
trackers_as_actions = []
# from multiple states that create equal featurizations
# we only need to keep one.
hashed_examples = set()
logger.debug(
"Creating states and action examples from "
"collected trackers (by {}({}))..."
"".format(type(self).__name__, type(self.state_featurizer).__name__)
)
pbar = tqdm(
trackers,
desc="Processed trackers",
disable=common_utils.is_logging_disabled(),
)
for tracker in pbar:
states = self._create_states(tracker, domain)
states_length_for_action = 0
for event in tracker.applied_events():
if not isinstance(event, ActionExecuted):
continue
states_length_for_action += 1
# use only actions which can be predicted at a stories start
if event.unpredictable:
continue
sliced_states = self.slice_state_history(
states[:states_length_for_action], self.max_history
)
if self.remove_duplicates:
hashed = self._hash_example(
sliced_states, event.action_name or event.action_text, tracker
)
# only continue with tracker_states that created a
# hashed_featurization we haven't observed
if hashed not in hashed_examples:
hashed_examples.add(hashed)
trackers_as_states.append(sliced_states)
trackers_as_actions.append(
[event.action_name or event.action_text]
)
else:
trackers_as_states.append(sliced_states)
trackers_as_actions.append([event.action_name or event.action_text])
pbar.set_postfix({"# actions": "{:d}".format(len(trackers_as_actions))})
logger.debug("Created {} action examples.".format(len(trackers_as_actions)))
return trackers_as_states, trackers_as_actions
def train(self, training_data, config, **kwargs):
# type: (TrainingData, Optional[rasaNLUModelConfig], **Any) -> None
"""Train the embedding intent classifier on a data set."""
intent_dict = self._create_intent_dict(training_data)
if len(intent_dict) < 2:
logger.error("Can not train an intent classifier. "
"Need at least 2 different classes. "
"Skipping training of intent classifier.")
return
self.inv_intent_dict = {v: k for k, v in intent_dict.items()}
self.encoded_all_intents = self._create_encoded_intents(intent_dict)
X, Y, intents_for_X = self._prepare_data_for_training(
training_data, intent_dict)
num_classes = len(intent_dict)
self.graph = tf.Graph()
with self.graph.as_default():
self.a_in = tf.placeholder(tf.float32, (None, X.shape[-1]),
name='a')
self.b_in = tf.placeholder(tf.float32, (None, Y.shape[-1]),
name='b')
is_training = tf.placeholder_with_default(False, shape=())
self.drop_out = tf.placeholder(tf.float32, (), name='drop_out')
# Create a graph for training
logits_train = conv_net(self.a_in,
num_classes,
self.num_hidden_layers,
self.hidden_layer_size,
self.C2,
self.drop_out,
is_training=True)
# Define loss and optimizer
# (with train logits, for dropout to take effect)
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(
logits=logits_train,
labels=self.b_in)) + tf.losses.get_regularization_loss()
self.y_predict = tf.nn.softmax(logits_train)
train_op = tf.train.AdamOptimizer(
learning_rate=self.learning_rate).minimize(loss)
# train tensorflow graph
config_proto = self.get_config_proto(self.component_config)
self.session = tf.Session(graph=self.graph, config=config_proto)
self.session.run(tf.global_variables_initializer())
pbar = tqdm(range(self.epochs),
desc="Epochs",
disable=is_logging_disabled())
train_acc = 0
last_loss = 0
for ep in pbar:
indices = np.random.permutation(len(X))
batch_size = self.batch_size
batches_per_epoch = (len(X) // batch_size +
int(len(X) % batch_size > 0))
ep_loss = 0
for i in range(batches_per_epoch):
end_idx = (i + 1) * batch_size
start_idx = i * batch_size
batch_a = X[indices[start_idx:end_idx]]
batch_b = Y[indices[start_idx:end_idx]]
sess_out = self.session.run(
{
'loss': loss,
'train_op': train_op
},
feed_dict={
self.a_in: batch_a,
self.b_in: batch_b,
is_training: True,
self.drop_out: self.droprate
})
ep_loss += sess_out.get('loss') / batches_per_epoch
if self.evaluate_on_num_examples:
if (ep == 0
or (ep + 1) % self.evaluate_every_num_epochs == 0
or (ep + 1) == self.epochs):
train_acc = self._output_training_stat(
X, intents_for_X, is_training)
last_loss = ep_loss
pbar.set_postfix({
"loss": "{:.3f}".format(ep_loss),
"acc": "{:.3f}".format(train_acc)
})
else:
pbar.set_postfix({"loss": "{:.3f}".format(ep_loss)})