def _episode_to_range(episode_number) -> Tuple[int, int]:
num_sentences = Sentence.objects.count()
if num_sentences == 0:
raise NoDataException
start = episode_number * Configuration.get_active().episode_size
end = (episode_number + 1) * Configuration.get_active().episode_size
if episode_number < 0:
start += num_sentences
end += num_sentences
return start, end
def on_message(self, sentence):
assert isinstance(sentence, Sentence)
logger.debug(
"Received notification on new message to bot, processing...")
factory = self._get_and_update_factory(sentence)
context = self._create_context_from_sentence(sentence)
state = State(sentence)
# mapping unknown_intent -> "Sorry didn't understand that" is hard coded, to avoid confusion
if sentence.intent.template.name != Configuration.get_active(
).unknown_intent.name:
action_name = self.bot.query({
'state_input':
numpy.array([state.as_vector()]),
'context_input':
numpy.array([context.as_matrix()])
})
else:
action_name = Configuration.get_active(
).didnt_understand_intent.name
human_user = User.objects.get(username=sentence.said_by)
chat = Chat.objects.get(initiator=human_user, receiver=self._bot_user)
sentence_value = factory.create_response(action_name)
message = Message.objects.create(value=sentence_value,
sent_by=self._bot_user,
sent_in=chat)
dialogue = sentence.said_in
intent_template = IntentTemplate.objects.get(name=action_name)
intent = Intent.objects.create(template=intent_template)
response = Sentence.objects.create(value=sentence_value,
said_in=dialogue,
said_by=self._bot_user,
raw_sentence=message,
intent=intent,
sentiment=0,
created_by=self.bot.model_base_name)
update_user_profile_for_single_dialogue(response.said_in)
response.refresh_from_db()
logger.info(
'Responded to sentence {}(intent="{}") with action "{}". Done processing message!'
.format(
sentence, sentence.intent.template.name if sentence.intent
is not None else 'Unknown intent', action_name))
self.bot.train()
def all_transitions_from_turns(
turns: List[Turn],
context_length=Configuration.get_active().context_length):
"""
Creates transitions from a list of turns. Reverses that list of turns, so put in a list of turns
with ascending "dates" of these turns.
:param turns: List of Turns
:param context_length: length of context for each turn
:return: a list of Transitions
"""
transitions = []
num_turns = len(turns)
turns = turns[::-1]
# one slice needs to have context_length turns for the context
# and one for each State s_0 (current state) as well as State s_1 (future state)
slice_size = context_length + 2
for i, turn in enumerate(turns):
remaining_turns = min(num_turns - i, slice_size)
turns_slice = turns[i:i + remaining_turns]
if len(turns_slice) >= 2:
# need at least 2 states for a proper transition
transition = Transition.single_transition_from_turns(
turns_slice, context_length)
transitions.append(transition)
return transitions
def vector_from_sentence(sentence: Sentence):
if sentence.intent is None:
raise NoIntentError(sentence)
if not Configuration.get_active().is_action_intent(
sentence.intent.template.name):
raise NoActionIntentError(sentence.intent, sentence)
return Action.vector_from_name(sentence.intent.template.name)
def __init__(self):
self._base_sentences = Configuration.get_active(
).response_templates.all()
for template in self._base_sentences:
template.prepare()
self._context = {}
self._tracking_table = {
'query.player.height': ['player'],
'query.player.information.age': ['player'],
'query.player.information.goals': ['player'],
'query.player.information.shoe': ['player'],
'query.player.news': ['content-type', 'player'],
'query.player.news.more': ['player', 'content-type'],
'userprofile.response.name': ['given-name'],
'userprofile.response.favorite_player': ['player'],
'userprofile.response.age': ['age']
}
self._response_builders = {
'common.hi': self._greeting,
'common.thanks': self._thanks,
'common.how_are_you': self._how_are_you,
'response.player.news': self._content,
'response.player.information.age': self._player_age,
'response.player.information.height': self._player_height,
'response.player.information.goals': self._player_goals,
'response.player.information.shoe': self._player_shoe,
'offer.player.news': self._offer_content,
}
self._content_interface = SimpleContentInterface()
def _init_callbacks(self):
"""
Initializes callbacks needed for this model.
Implement to specify callbacks used.
Defaults to using only TensorBoard callback.
"""
self._callbacks = [
TensorBoard(log_dir=os.path.join(
Configuration.get_active().log_dir, self._model.name))
]
def get_single_context(turns, context_length: int = Configuration.get_active().context_length):
"""
Creates a context object from turns, useful for training
:param turns: the Turns that make up the context, e.g. the turns prior to the current Sentence/State
:param context_length: number of steps into the past
:return: a Context object see __init__
"""
assert len(turns) <= context_length
states = [turn.user for turn in turns]
actions = [turn.bot for turn in turns]
return Context(states, actions, context_length)
def __init__(self, sentence: Sentence = None):
"""
creates a new state, if sentence is given, the fields will be populated with values from this sentence
:param sentence: an object of type Sentence, which will be used to populate the state
"""
if sentence is not None:
assert isinstance(sentence, Sentence)
self.intent_name = sentence.intent.template.name
self.intent_vector = State._intent_vector_from_sentence(sentence)
try:
self._intent_index = Configuration.get_active().state_index_for_name(self.intent_name)
except ValueError:
# not in list --> unknown intent
self._intent_index = Configuration.get_active().state_index_for_name(
Configuration.get_active().unknown_intent.name
)
self.sentiment = float(sentence.sentiment)
self.user_profile = sentence.user_profile
self.user_profile_vector = State._convert_user_profile(sentence.user_profile)
def predict(self, inputs: Dict[str, numpy.ndarray]):
"""
Simple delegate for the keras.model.predict function
:param inputs: input as dict of input_name->numpy.ndarray
:return: an array of quality vectors for each input row (state/context pair), not processed,
see keras.model.predict
"""
with self._graph.as_default():
return self._model.predict(
inputs, batch_size=Configuration.get_active().batch_size)
def single_transition_from_turns(
turns: List[Turn],
context_length=Configuration.get_active().context_length):
assert len(
turns
) >= 2, 'You need at least 2 turns for a Transition (= s_0->s_1 with no context)'
final_turn = turns[-1]
current_turn = turns[-2]
context = Context.get_single_context(turns[:-2], context_length)
transition = Transition(current_turn.user, current_turn.bot, context,
final_turn.user)
return transition
def __init__(self, sentence: Sentence = None):
"""
:param sentence: sentence to get the action from
:raises NoIntentError: if the given sentence has no intent
:raises NoActionIntentError: if the given sentence has an intent, that cannot be interpreted as action
"""
self._action_vector = Action.vector_from_sentence(sentence)
self._action_index = Configuration.get_active().action_index_for_name(
sentence.intent.template.name)
self.reward = float(sentence.reward)
self.terminal = bool(sentence.terminal)
def run(self):
# TODO: why?
sleep(5)
logger.info("Started TurnsTerminator.")
logger.info("TurnsTerminator checking for unterminated sentences.")
while True:
last_sentence_in_dialogues = Sentence.objects.raw(
'SELECT * FROM turns_sentence GROUP BY said_in_id ORDER BY said_on DESC;'
)
for sentence in last_sentence_in_dialogues:
pause = timezone.now() - sentence.said_on
pause = pause.seconds + pause.days * Configuration.get_active(
).seconds_per_day
if pause >= Configuration.get_active(
).seconds_for_terminal and not sentence.terminal:
logger.info(
"Got no new sentence after {} seconds, {} has to be a terminal sentence"
.format(pause, sentence.value))
sentence.terminal = True
sentence.save()
# Nyquist frequency, so we don't miss terminals
sleep(Configuration.get_active().seconds_for_terminal / 2)
def train(self, validate=True):
"""Trains for a single episode
:return: a bool indicating whether or not training was possible and successful.
"""
if not self.training_allowed:
return False
logger.debug('Acquiring lock...')
with self._training_lock:
if not self._can_sample_batch(self.episodes_seen):
logger.debug('No more episodes to train on are available!')
return False
with self._graph.as_default():
episode_logs = {}
training_loss = 0
batches_trained = 0
logger.info(
'Bot training on episode #{}'.format(self.episodes_seen +
1))
for step in range(
0,
Configuration.get_active().steps_per_episode):
num_samples, x, y = self._sample_batch(self.episodes_seen)
if num_samples > 0:
loss = self._perform_training_step((x, y))
batch_logs = {'loss': loss}
self._on_batch_end(batches_trained, batch_logs)
training_loss += batch_logs['loss']
batches_trained += 1
self.samples_seen += num_samples
else:
return False
training_loss /= batches_trained
episode_logs['loss'] = training_loss
if validate:
num_samples, x, y = self._sample_batch(self.episodes_seen)
if num_samples > 0:
episode_logs['val_loss'] = self._model.test_on_batch(
x, y)
episode_logs[
'est_value'] = self._model.predict_on_batch(x)
episode_logs['est_value'] = episode_logs[
'est_value'].sum() / num_samples
episode_logs['true_value'] = y['quality_output'].sum(
) / num_samples
self.episodes_seen += 1
self._save_stats()
self._save_weights()
self._on_episode_end(self.episodes_seen, episode_logs)
return True
def __str__(self) -> str:
return '<Action {}>'.format(
Configuration.get_active().action_intents[self._action_index].name)
def __init__(self):
super().__init__(float(Configuration.get_active().epsilon))
def _sample_batch(self, episode_number) -> Tuple[int, Dict, Dict]:
start, stop = DeepMindNoContextBot._episode_to_range(episode_number)
sentences = \
[Sentence.sample_sentence_in_range(self._bot_user.username, start, stop) for _ in
range(0, Configuration.get_active().batch_size)]
transitions = []
for sentence in sentences:
context_sentences = []
while len(context_sentences) < 4:
assert sentence.said_by == self._bot_user.username
context_sentences = Sentence.objects.filter(
said_in=sentence.said_in).filter(
said_on__lte=sentence.said_on).order_by(
'-said_on'
)[:Configuration.get_active().context_length * 2 + 4]
context_sentences = list(reversed(context_sentences))
if len(context_sentences) < 4:
# re-sample, since this sample has not enough context
sentence = Sentence.sample_sentence_in_range(
self._bot_user.username, start, stop)
try:
a1 = Action(context_sentences.pop())
s1 = State(context_sentences.pop())
turns = Turn.sentences_to_turns(context_sentences,
self._bot_user)
# context turns list should only be CONTEXT_LENGTH long
raw_context_t1 = turns[0:Configuration.get_active().
context_length]
context_t1 = Context.get_single_context(
raw_context_t1,
Configuration.get_active().context_length)
a0 = Action(context_sentences.pop())
s0 = State(context_sentences.pop())
turns = Turn.sentences_to_turns(context_sentences,
self._bot_user)
context_t0 = Context.get_single_context(
turns,
Configuration.get_active().context_length)
except IntentError as e:
logger.error(
'Error occurred while processing sampled sentence {}. See below.'
.format(sentence))
raise e
transition = Transition(s0, a0, context_t0, s1, a1, context_t1)
transitions.append(transition)
if len(transitions) < Configuration.get_active().batch_size:
return 0, {}, {}
assert len(transitions) == Configuration.get_active().batch_size
actions = [transition.action_t0 for transition in transitions]
states = numpy.array(
[transition.state_t0.as_vector() for transition in transitions])
contexts = numpy.array(
[transition.context_t0.as_matrix() for transition in transitions])
future_states = numpy.array(
[transition.state_t1.as_vector() for transition in transitions])
future_contexts = numpy.array(
[transition.context_t1.as_matrix() for transition in transitions])
terminals = numpy.array(
[0 if transition.terminal else 1 for transition in transitions])
rewards = numpy.array(
[transition.action_t0.reward for transition in transitions])
assert future_states.shape == (Configuration.get_active(
).batch_size, ) + Configuration.get_active().state_shape
assert contexts.shape == (Configuration.get_active().batch_size,
) + Configuration.get_active().context_shape
assert rewards.shape == (Configuration.get_active().batch_size, )
target_quality = self._target.predict({
'state_input': future_states,
'context_input': future_contexts
})
assert target_quality.shape == (
Configuration.get_active().batch_size,
Configuration.get_active().number_actions)
quality_batch = target_quality.max(axis=1).flatten()
quality_batch *= self.discount
quality_batch *= terminals
logger.debug("Future qualities are {}".format(quality_batch))
logger.debug('Rewards are {}'.format(rewards))
quality_batch = rewards + quality_batch
logger.debug("Working with qualities {}".format(quality_batch))
target_quality = numpy.zeros(
(Configuration.get_active().batch_size,
Configuration.get_active().number_actions))
for target, action, quality in zip(target_quality, actions,
quality_batch):
target[action.intent_index] = quality
return (len(states), {
'state_input': states,
'context_input': contexts
}, {
'quality_output': target_quality
})
def model_directory(self):
return os.path.join(Configuration.get_active().weights_dir,
self._model.name)
def __init__(self):
super().__init__(float(Configuration.get_active().discount))
def _prediction_to_action_name(self, prediction: numpy.ndarray) -> str:
prediction = prediction[0]
return list(Configuration.get_active().action_intents.all())[
prediction.argmax()].name