def __init__(self):
import os
#实体追踪
et = EntityTracker()
#词袋 word2vec
self.bow_enc = BoW_encoder()
#加载word2vec embedding
self.emb = UtteranceEmbed()
#将实体追踪器添加到动作追踪器中
at = ActionTracker(et)
#得到数据集和对话开始 结束行数
self.dataset, dialog_indices = Data(et, at).trainset
#划分数据集：200对做训练 50对做测试
self.dialog_indices_tr = dialog_indices
self.dialog_indices_dev = dialog_indices
#obs_size 300维的词向量 + 85个袋中的词 + 4个槽位
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
#话术模板
self.action_templates = at.get_action_templates()
#动作个数
action_size = at.action_size
#隐藏层神经元个数
nb_hidden = 128
self.net = LSTM_net(obs_size=obs_size,
action_size=action_size,
nb_hidden=nb_hidden)
def __init__(self):
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
self.at = ActionTracker(interact=True)
print(self.at.dialog_acts)
_, _, self.other_da, self.other_da_id = self.at.get_data()
self.action_templates = self.at.action_templates
self.action_size = self.at.action_size
#self.responses = [r[0] for r in self.at.responses]
#print(self.emb.dim , self.bow_enc.vocab_size , self.action_size)
self.obs_size = self.emb.dim + self.bow_enc.vocab_size + self.action_size
nb_hidden = 128
self.exercises = list(self.at.action_templates.keys())
self.net = LSTM_net(obs_size=self.obs_size,
action_size=self.action_size,
nb_hidden=nb_hidden)
# restore checkpoint
self.net.restore()
# setup remote interaction via zeromq
context = zmq.Context.instance()
# PUB socket for sending system-output
self.pubSocket: zmq.Socket = context.socket(zmq.PUB)
self.pubSocket.bind(Config.SYSTEM_OUTPUT_ADDR)
# SUB socket for receiving user-input
self.subSocket: zmq.Socket = context.socket(zmq.SUB)
self.subSocket.connect(Config.USER_INPUT_ADDR)
self.subSocket.setsockopt_string(zmq.SUBSCRIBE,
Config.USER_INPUT_SUBJECT)
self.lastOutputMessage: str = None
self.requestHandler: threading.Thread = None
self.requestInterrupted: threading.Lock = None
def __init__(self):
et = EntityTracker()
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
at = ActionTracker(et)
self.train_dataset, train_dialog_indices = Data(et, at).train_set
self.test_dataset, test_dialog_indices = Data(et, at).test_set
print('=========================\n')
print('length of Train dialog indices : ', len(train_dialog_indices))
print('=========================\n')
print('=========================\n')
print('length of Test dialog indices : ', len(test_dialog_indices))
print('=========================\n')
# Shuffle Training Dataset
random.shuffle(train_dialog_indices)
self.dialog_indices_tr = train_dialog_indices
self.dialog_indices_dev = test_dialog_indices
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
self.action_templates = at.get_action_templates()
action_size = at.action_size
# nb_hidden = 128
nb_hidden = 150
print('=========================\n')
print('Action_templates: ', action_size)
print('=========================\n')
self.net = LSTM_net(obs_size=obs_size,
action_size=action_size,
nb_hidden=nb_hidden)
self.et = et
self.at = at
action_projection = []
for action in self.action_templates:
action_projection.append(self.emb.encode(action))
self.action_projection = np.transpose(action_projection)
self.action_size = action_size
def __init__(self):
et = EntityTracker()
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
at = ActionTracker(et)
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
self.action_templates = at.get_action_templates()
action_size = at.action_size
nb_hidden = 128
self.net = LSTM_net(obs_size=obs_size, action_size=action_size, nb_hidden=nb_hidden)
# restore checkpoint
self.net.restore()
def __init__(self):
et = EntityTracker()
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
at = ActionTracker(et)
self.dataset, dialog_indices = Data(et, at).trainset
self.dialog_indices_tr = dialog_indices[:200]
self.dialog_indices_dev = dialog_indices[200:250]
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
self.action_templates = at.get_action_templates()
action_size = at.action_size
nb_hidden = 128
self.net = LSTM_net(obs_size=obs_size,
action_size=action_size,
nb_hidden=nb_hidden)
def __init__(self, train_ratio=15/20, \
epochs=100, \
train_whole=True, \
):
self.train_ratio = train_ratio
self.epochs = epochs
#flag für fehleranalyse
self.startCountingErrors = False
self.count_all_one_eval = 0
self.count_all_one_train = 0
self.count_all_train = 0
self.count_all_eval = 0
#data.trainset = [(u,da_indx), ...]
#HIER acttemps und size
self.at = ActionTracker()
self.action_templates = self.at.action_templates
self.at.dialog_acts = self.at.dialog_acts
self.at.dialog_act_ids = self.at.dialog_act_ids
self.dataset, self.dialog_indices, self.other_da, self.other_da_id = self.at.get_data()
self.action_size = self.at.action_size
self.accs = []
self.loss = []
self.dev_accs =[]
self.dev_loss=[]
self.countErrors = Counter()
self.countPredicitions = 0
self.errorLog = {}
self.highest_accuracy = 0
#pickle.load(open("data/errors/errorLog.p", "rb"))
if train_whole:
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
#HIER numfeats
obs_size = self.emb.dim + self.bow_enc.vocab_size + self.action_size# + self.et.size_context_features
nb_hidden = 128
self.net = LSTM_net(obs_size,
self.action_size,
nb_hidden=nb_hidden)
def __init__(self):
et = EntityTracker()
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
at = ActionTracker(et)
'''
['any preference on a type of cuisine', 'api_call <cuisine> <location> <party_size> <rest_type>', 'great let me do the reservation', 'hello what can i help you with today', 'here it is <info_address>', 'here it is <info_phone>', 'how many people would be in your party', "i'm on it", 'is there anything i can help you with', 'ok let me look into some options for you', 'sure is there anything else to update', 'sure let me find an other option for you', 'what do you think of this option: <restaurant>', 'where should it be', 'which price range are looking for', "you're welcome"]
'''
self.dataset, dialog_indices = Data(et, at).trainset
self.dialog_indices_tr = dialog_indices[:200]
self.dialog_indices_dev = dialog_indices[200:250]
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
self.action_templates = at.get_action_templates()
action_size = at.action_size
nb_hidden = 128
self.net = LSTM_net(obs_size=obs_size,
action_size=action_size,
nb_hidden=nb_hidden)
def __init__(self):
et = EntityTracker()
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
at = ActionTracker(et)
self.dataset, dialog_indices = Data(et, at).trainset
train_indices = joblib.load('data/train_test_list/train_indices_759')
test_indices = joblib.load('data/train_test_list/test_indices_759_949')
self.dialog_indices_tr = train_indices
self.dialog_indices_dev = test_indices
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
self.action_templates = at.get_action_templates()
action_size = at.action_size
nb_hidden = 128
self.net = LSTM_net(obs_size=obs_size,
action_size=action_size,
nb_hidden=nb_hidden)
class InteractiveSession:
def __init__(self):
et = EntityTracker()
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
at = ActionTracker(et)
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
self.action_templates = at.get_action_templates()
action_size = at.action_size
nb_hidden = 128
self.net = LSTM_net(obs_size=obs_size, action_size=action_size, nb_hidden=nb_hidden)
# restore checkpoint
self.net.restore()
def interact(self):
# create entity tracker
et = EntityTracker()
# create action tracker
at = ActionTracker(et)
# reset network
self.net.reset_state()
# begin interaction loop
while True:
# get input from user
u = input('User: '******'clear' or u == 'reset' or u == 'restart':
self.net.reset_state()
et = EntityTracker()
at = ActionTracker(et)
print('Bot: Reset successfully')
# check for entrance and exit command
elif u == 'exit' or u == 'stop' or u == 'quit' or u == 'q':
print("Bot: Thank you for using")
break
elif u == 'hello' or u == 'hi':
print("Bot: Hello, what can i do for you")
elif u == 'thank you' or u == 'thanks' or u == 'thank you very much':
print('Bot: You are welcome')
break
else:
if not u:
continue
u = u.lower()
# encode
u_ent = et.extract_entities(u)
u_ent_features = et.context_features() # 5
# print(et.entities)
# print(et.ctxt_features)
u_emb = self.emb.encode(u) # 300
u_bow = self.bow_enc.encode(u) # 60
# concat features
features = np.concatenate((u_ent_features, u_emb, u_bow), axis=0)
# print(features.shape)
# get action mask
action_mask = at.action_mask()
# action_mask = np.ones(self.net.action_size)
# print("action_mask: ", action_mask)
# forward
prediction = self.net.forward(features, action_mask)
response = self.action_templates[prediction]
if prediction == 0:
slot_values = copy.deepcopy(et.entities)
slot_values.pop('<location>')
memory = []
count = 0
for k, v in slot_values.items():
memory.append('='.join([k, v]))
count += 1
if count == 2:
memory.append('\n')
response = response.replace("memory", ', '.join(memory))
# memory = ', '.join(slot_values.values())
# response = response.replace("memory", memory)
self.net.reset_state()
et = EntityTracker()
at = ActionTracker(et)
# print('Execute successfully and begin new session')
if prediction == 1:
response = response.replace("location", '<location>=' + et.entities['<location>'])
print('Bot: ', response)
class InteractiveSession():
def __init__(self):
et = EntityTracker()
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
at = ActionTracker(et)
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
self.action_templates = at.get_action_templates()
action_size = at.action_size
nb_hidden = 128
self.net = LSTM_net(obs_size=obs_size,
action_size=action_size,
nb_hidden=nb_hidden)
# restore checkpoint
self.net.restore()
def interact(self):
# create entity tracker
et = EntityTracker()
# create action tracker
at = ActionTracker(et)
# reset network
self.net.reset_state()
# begin interaction loop
while True:
# get input from user
u = input(':: ')
# check if user wants to begin new session
if u == 'clear' or u == 'reset' or u == 'restart':
self.net.reset_state()
et = EntityTracker()
at = ActionTracker(et)
print('')
# check for exit command
elif u == 'exit' or u == 'stop' or u == 'quit' or u == 'q':
break
else:
# ENTER press : silence
if not u:
u = '<SILENCE>'
# encode
u_ent, u_entities = et.extract_entities(u, is_test=True)
u_ent_features = et.context_features()
u_emb = self.emb.encode(u)
u_bow = self.bow_enc.encode(u)
# concat features
features = np.concatenate((u_ent_features, u_emb, u_bow),
axis=0)
# get action mask
action_mask = at.action_mask()
# forward
prediction = self.net.forward(features, action_mask)
print('prediction : ', prediction)
print(u_entities)
print('\n')
if self.post_process(prediction, u_ent_features):
print(
'>>', 'api_call ' + u_entities['<cuisine>'] + ' ' +
u_entities['<location>'] + ' ' +
u_entities['<party_size>'] + ' ' +
u_entities['<rest_type>'])
else:
prediction = self.action_post_process(
prediction, u_entities)
print('>>', self.action_templates[prediction])
# if all entities is satisfied and the user agree to make a reservation.
if all(u_ent_featur == 1
for u_ent_featur in u_ent_features) and (prediction
== 10):
break
def post_process(self, prediction, u_ent_features):
if prediction == 0:
return True
attr_list = [9, 12, 6, 1]
if all(u_ent_featur == 1
for u_ent_featur in u_ent_features) and prediction in attr_list:
return True
else:
return False
def action_post_process(self, prediction, u_entities):
attr_mapping_dict = {
9: '<cuisine>',
12: '<location>',
6: '<party_size>',
1: '<rest_type>'
}
# find exist and non-exist entity
exist_ent_index = [
key for key, value in u_entities.items() if value != None
]
non_exist_ent_index = [
key for key, value in u_entities.items() if value == None
]
# if predicted key is already in exist entity index then find non exist entity index
# and leads the user to input non exist entity.
if prediction in attr_mapping_dict:
pred_key = attr_mapping_dict[prediction]
if pred_key in exist_ent_index:
for key, value in attr_mapping_dict.items():
if value == non_exist_ent_index[0]:
return key
else:
return prediction
else:
return prediction
class InteractiveSession():
def __init__(self):
et = EntityTracker()
self.bow_enc = BoW_encoder(et)
self.emb = UtteranceEmbed()
at = ActionTracker(et)
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
self.action_templates = at.get_action_templates()
action_size = at.action_size
nb_hidden = 128
self.net = LSTM_net(obs_size=obs_size,
action_size=action_size,
nb_hidden=nb_hidden)
# restore checkpoint
self.net.restore()
def interact(self, input):
# create entity tracker
et = EntityTracker()
# create action tracker
at = ActionTracker(et)
# reset network
self.net.reset_state()
# begin interaction loop
#while True:
# get input from user
#u = input(':: ')
u = input
# check if user wants to begin new session
if u == 'clear' or u == 'reset' or u == 'restart':
self.net.reset_state()
et = EntityTracker()
at = ActionTracker(et)
print('')
# check for exit command
#elif u == 'exit' or u == 'stop' or u == 'quit' or u == 'q':
#break
else:
# ENTER press : silence
if not u:
u = '<SILENCE>'
# encode
u_ent = et.extract_entities(u)
u_ent_features = et.context_features()
u_emb = self.emb.encode(u)
u_bow = self.bow_enc.encode(u)
# concat features
features = np.concatenate((u_ent_features, u_emb, u_bow), axis=0)
# get action mask
action_mask = at.action_mask()
# forward
prediction = self.net.forward(features, action_mask)
#print('>>', self.action_templates[prediction])
return self.action_templates[prediction]
class Dialogue():
def __init__(self):
# stor whole dialogues
self.story = []
self.sp_confidecne = []
self.file_path = os.path.join(
rospkg.RosPack().get_path('dialogue_system'), 'log',
'dialogue.txt')
# count turn taking
self.usr_count = 0
self.sys_count = 0
# paramaters
self.network_type = rospy.get_param('~network_model', 'stacked_lstm')
self.lang_type = rospy.get_param('~lang', 'eng')
self.is_emb = rospy.get_param('~embedding', 'false')
self.is_am = rospy.get_param('~action_mask', "true")
self.user_num = rospy.get_param('~user_number', '0')
# call rest of modules
self.et = EntityTracker()
self.at = ActionTracker(self.et)
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed(lang=self.lang_type)
# select observation size for RNN
if self.is_am and self.is_emb:
obs_size = self.emb.dim + self.bow_enc.vocab_size + self.et.num_features + self.at.action_size
elif self.is_am and not (self.is_emb):
obs_size = self.bow_enc.vocab_size + self.et.num_features + self.at.action_size
elif not (self.is_am) and self.is_emb:
obs_size = self.emb.dim + self.bow_enc.vocab_size + self.et.num_features
elif not (self.is_am) and not (self.is_emb):
obs_size = self.bow_enc.vocab_size + self.et.num_features
self.action_template = self.at.get_action_templates()
self.at.do_display_template()
# must clear entities space
self.et.do_clear_entities()
action_size = self.at.action_size
nb_hidden = 128
if self.network_type == 'gru':
self.net = GRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
elif self.network_type == 'reversed_lstm':
self.net = ReversingLSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
elif self.network_type == 'reversed_gru':
self.net = ReversingGRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
elif self.network_type == 'stacked_gru':
self.net = StackedGRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
elif self.network_type == 'stacked_lstm':
self.net = StackedLSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
elif self.network_type == 'lstm':
self.net = LSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
elif self.network_type == 'bidirectional_lstm':
self.net = BidirectionalLSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
elif self.network_type == 'bidirectional_gru':
self.net = BidirectionalGRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
# restore trained model
self.net.restore()
# rostopics
self.pub_reply = rospy.Publisher('reply', Reply, queue_size=10)
self.pub_complete = rospy.Publisher('complete_execute_scenario',
Empty,
queue_size=10)
rospy.Subscriber('raising_events', RaisingEvents,
self.handle_raise_events)
try:
rospy.wait_for_service('reception_db/query_data')
self.get_response_db = rospy.ServiceProxy(
'reception_db/query_data', DBQuery)
rospy.logwarn("waiting for reception DB module...")
except rospy.exceptions.ROSInterruptException as e:
rospy.logerr(e)
quit()
rospy.logwarn(
"network: {}, lang: {}, action_mask: {}, embedding: {}, user_number: {}"
.format(self.network_type, self.lang_type, self.is_am, self.is_emb,
self.user_num))
self.story.append('user number: %s' % self.user_num)
rospy.loginfo('\033[94m[%s]\033[0m initialized.' % rospy.get_name())
# if utterance == 'clear':
# self.net.reset_state()
# self.et.do_clear_entities()
# response = 'context has been cleared.'
def get_response(self, utterance):
rospy.loginfo("actual input: %s" %
utterance) # check actual user input
# clean utterance
# utterance = re.sub(r'[^ a-z A-Z 0-9]', " ", utterance)
# utterance preprocessing
u_ent, u_entities = self.et.extract_entities(utterance, is_test=True)
u_ent_features = self.et.context_features()
u_bow = self.bow_enc.encode(utterance)
if self.is_emb:
u_emb = self.emb.encode(utterance)
try:
if self.is_am:
action_mask = self.at.action_mask()
# concatenated features
if self.is_am and self.is_emb:
features = np.concatenate(
(u_ent_features, u_emb, u_bow, action_mask), axis=0)
elif self.is_am and not (self.is_emb):
features = np.concatenate((u_ent_features, u_bow, action_mask),
axis=0)
elif not (self.is_am) and self.is_emb:
features = np.concatenate((u_ent_features, u_emb, u_bow),
axis=0)
elif not (self.is_am) and not (self.is_emb):
features = np.concatenate((u_ent_features, u_bow), axis=0)
# try:
# predict template number
if self.is_am:
probs, prediction = self.net.forward(features, action_mask)
else:
probs, prediction = self.net.forward(features)
# check response confidence
if max(probs) > BOUNDARY_CONFIDENCE:
response = self.action_template[prediction]
prediction = self.pre_action_process(prediction, u_entities)
# handle api call
if self.post_process(prediction, u_entities):
if prediction == 1:
response = 'api_call appointment {} {} {} {} {} {} {}'.format(
u_entities['<first_name>'],
u_entities['<last_name>'],
u_entities['<address_number>'],
u_entities['<address_name>'],
u_entities['<address_type>'], u_entities['<time>'],
u_entities['<pm_am>'])
elif prediction == 2:
response = 'api_call location {}'.format(
u_entities['<location>'])
elif prediction == 3:
response = 'api_call prescription {} {} {} {} {}'.format(
u_entities['<first_name>'],
u_entities['<last_name>'],
u_entities['<address_number>'],
u_entities['<address_name>'],
u_entities['<address_type>'])
elif prediction == 4:
response = 'api_call waiting_time {} {} {} {} {} {} {}'.format(
u_entities['<first_name>'],
u_entities['<last_name>'],
u_entities['<address_number>'],
u_entities['<address_name>'],
u_entities['<address_type>'], u_entities['<time>'],
u_entities['<pm_am>'])
response = self.get_response_db(
response
) # query knowledge base; here we use dynamo db
response = response.response
elif prediction in [6, 9, 11]:
response = self.action_template[prediction]
response = response.split(' ')
response = [
word.replace('<first_name>',
u_entities['<first_name>'])
for word in response
]
response = ' '.join(response)
else:
response = self.action_template[prediction]
else:
response = random.choice(
REPROMPT
) # if prediction confidence less than 40%, reprompt
except:
response = random.choice(REPROMPT)
return prediction, probs, response
def handle_raise_events(self, msg):
utterance = msg.recognized_word
try:
# get confidence
data = json.loads(msg.data[0])
confidence = data['confidence']
except:
confidence = None
if confidence > BOUNDARY_CONFIDENCE or confidence == None:
if 'silency_detected' in msg.events:
utterance = '<SILENCE>'
else:
try:
self.story.append(
"U%i: %s (sp_conf:%f)" %
(self.usr_count + 1, utterance, confidence))
self.sp_confidecne.append(confidence)
except:
self.story.append("U%i: %s" %
(self.usr_count + 1, utterance))
self.usr_count += 1
utterance = utterance.lower()
# generate system response
prediction, probs, response = self.get_response(utterance)
else:
prediction = -1
probs = -1
response = random.choice(REPROMPT)
# add system turn
self.story.append("A%i: %s" % (self.sys_count + 1, response))
self.sys_count += 1
# finish interaction
if (prediction == 6):
self.pub_complete.publish()
# logging user and system turn
self.story.append("user: %i, system: %i" %
(self.usr_count, self.sys_count))
self.story.append("mean_sp_conf: %f" %
(reduce(lambda x, y: x + y, self.sp_confidecne) /
len(self.sp_confidecne)))
self.story.append(
'==================================================================='
)
self.write_file(self.file_path, self.story)
# display system response
rospy.loginfo(json.dumps(self.et.entities,
indent=2)) # recognized entity values
try:
rospy.logwarn("System: [conf: %f, predict: %d] / %s\n" %
(max(probs), prediction, response))
except:
rospy.logwarn("System: [] / %s\n" % (response))
reply_msg = Reply()
reply_msg.header.stamp = rospy.Time.now()
reply_msg.reply = response
self.pub_reply.publish(reply_msg)
def post_process(self, prediction, u_ent_features):
api_call_list = [1, 2, 3, 4]
if prediction in api_call_list:
return True
attr_list = [0, 9, 10, 11, 12]
if all(u_ent_featur == 1
for u_ent_featur in u_ent_features) and prediction in attr_list:
return True
else:
return False
def action_post_process(self, prediction, u_entities):
attr_mapping_dict = {
11: '<first_name>',
11: '<last_name>',
12: '<address_number>',
12: '<address_name>',
12: '<address_type>',
10: '<time>',
10: '<pm_am>',
}
# find exist and non-exist entity
exist_ent_index = [
key for key, value in u_entities.items() if value != None
]
non_exist_ent_index = [
key for key, value in u_entities.items() if value == None
]
# if predicted key is already in exist entity index then find non exist entity index
# and leads the user to input non exist entity.
if prediction in attr_mapping_dict:
pred_key = attr_mapping_dict[prediction]
if pred_key in exist_ent_index:
for key, value in attr_mapping_dict.items():
if value == non_exist_ent_index[0]:
return key
else:
return prediction
else:
return prediction
def pre_action_process(self, prediction, u_entities):
api_call_list = [1, 3, 4]
attr_mapping_dict = {
'<first_name>': 11,
'<last_name>': 11,
'<address_number>': 12,
'<address_name>': 12,
'<address_type>': 12,
'<time>': 10,
'<pm_am>': 10,
}
# find exist and non-exist entity
non_exist_ent_index = [
key for key, value in u_entities.items() if value == None
]
if prediction in api_call_list:
if '<first_name>' in non_exist_ent_index:
prediction = attr_mapping_dict['<first_name>']
return prediction
'''
writing story log file
'''
def write_file(self, path, story_list):
with open(path, 'a') as f:
for item in story_list:
f.write("%s\n" % item)
rospy.logwarn('save dialogue histories.')
class Trainer():
def __init__(self):
import os
#实体追踪
et = EntityTracker()
#词袋 word2vec
self.bow_enc = BoW_encoder()
#加载word2vec embedding
self.emb = UtteranceEmbed()
#将实体追踪器添加到动作追踪器中
at = ActionTracker(et)
#得到数据集和对话开始 结束行数
self.dataset, dialog_indices = Data(et, at).trainset
#划分数据集：200对做训练 50对做测试
self.dialog_indices_tr = dialog_indices
self.dialog_indices_dev = dialog_indices
#obs_size 300维的词向量 + 85个袋中的词 + 4个槽位
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
#话术模板
self.action_templates = at.get_action_templates()
#动作个数
action_size = at.action_size
#隐藏层神经元个数
nb_hidden = 128
self.net = LSTM_net(obs_size=obs_size,
action_size=action_size,
nb_hidden=nb_hidden)
def train(self):
print('\n:: training started\n')
epochs = 25
for j in range(epochs):
# iterate through dialogs
#训练集个数
num_tr_examples = len(self.dialog_indices_tr)
loss = 0.
for i, dialog_idx in enumerate(self.dialog_indices_tr):
# get start and end index
start, end = dialog_idx['start'], dialog_idx['end']
# train on dialogue
loss += self.dialog_train(self.dataset[start:end])
# print #iteration
sys.stdout.write('\r{}.[{}/{}]'.format(j + 1, i + 1,
num_tr_examples))
print('\n\n:: {}.tr loss {}'.format(j + 1, loss / num_tr_examples))
# evaluate every epoch
accuracy = self.evaluate()
print(':: {}.dev accuracy {}\n'.format(j + 1, accuracy))
if accuracy > 0.4:
self.net.save()
continue
#训练过程
def dialog_train(self, dialog):
# create entity tracker
et = EntityTracker()
# create action tracker
at = ActionTracker(et)
# reset network
self.net.reset_state()
loss = 0.
# iterate through dialog
#u 用户输入 r 对应的动作索引
for (u, r) in dialog:
#u_ent 分词后的字符串
u_ent = et.extract_entities(u)
#槽位填充情况 【0 0 0 0】
u_ent_features = et.context_features()
#word2vec
u_emb = self.emb.encode(u)
#multi-hot
u_bow = self.bow_enc.encode(u)
# concat features
#300 + 85 + 4 = 389
features = np.concatenate((u_ent_features, u_emb, u_bow), axis=0)
# get action mask
action_mask = at.action_mask()
# forward propagation
# train step
loss += self.net.train_step(features, r, action_mask)
return loss / len(dialog)
#评估acc
def evaluate(self):
# create entity tracker
et = EntityTracker()
# create action tracker
at = ActionTracker(et)
# reset network
self.net.reset_state()
dialog_accuracy = 0.
#加载测试集
for dialog_idx in self.dialog_indices_dev:
start, end = dialog_idx['start'], dialog_idx['end']
dialog = self.dataset[start:end]
num_dev_examples = len(self.dialog_indices_dev)
# create entity tracker
et = EntityTracker()
# create action tracker
at = ActionTracker(et)
# reset network
self.net.reset_state()
# iterate through dialog
correct_examples = 0
#对于每个dialog 提取出utterance 和 response
for (u, r) in dialog:
# encode utterance
#提取出user中带有的实体
u_ent = et.extract_entities(u)
#提取当前槽位填充情况
u_ent_features = et.context_features()
u_emb = self.emb.encode(u)
u_bow = self.bow_enc.encode(u)
# concat features
features = np.concatenate((u_ent_features, u_emb, u_bow),
axis=0)
# get action mask 16维的multi-hot 向量
action_mask = at.action_mask()
# forward propagation
# train step
prediction = self.net.forward(features, action_mask)
correct_examples += int(prediction == r)
# get dialog accuracy
dialog_accuracy += correct_examples / len(dialog)
return dialog_accuracy / num_dev_examples
def __init__(self):
# stor whole dialogues
self.story = []
self.sp_confidecne = []
self.file_path = os.path.join(
rospkg.RosPack().get_path('dialogue_system'), 'log',
'dialogue.txt')
# count turn taking
self.usr_count = 0
self.sys_count = 0
# paramaters
self.network_type = rospy.get_param('~network_model', 'stacked_lstm')
self.lang_type = rospy.get_param('~lang', 'eng')
self.is_emb = rospy.get_param('~embedding', 'false')
self.is_am = rospy.get_param('~action_mask', "true")
self.user_num = rospy.get_param('~user_number', '0')
# call rest of modules
self.et = EntityTracker()
self.at = ActionTracker(self.et)
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed(lang=self.lang_type)
# select observation size for RNN
if self.is_am and self.is_emb:
obs_size = self.emb.dim + self.bow_enc.vocab_size + self.et.num_features + self.at.action_size
elif self.is_am and not (self.is_emb):
obs_size = self.bow_enc.vocab_size + self.et.num_features + self.at.action_size
elif not (self.is_am) and self.is_emb:
obs_size = self.emb.dim + self.bow_enc.vocab_size + self.et.num_features
elif not (self.is_am) and not (self.is_emb):
obs_size = self.bow_enc.vocab_size + self.et.num_features
self.action_template = self.at.get_action_templates()
self.at.do_display_template()
# must clear entities space
self.et.do_clear_entities()
action_size = self.at.action_size
nb_hidden = 128
if self.network_type == 'gru':
self.net = GRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
elif self.network_type == 'reversed_lstm':
self.net = ReversingLSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
elif self.network_type == 'reversed_gru':
self.net = ReversingGRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
elif self.network_type == 'stacked_gru':
self.net = StackedGRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
elif self.network_type == 'stacked_lstm':
self.net = StackedLSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
elif self.network_type == 'lstm':
self.net = LSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
elif self.network_type == 'bidirectional_lstm':
self.net = BidirectionalLSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
elif self.network_type == 'bidirectional_gru':
self.net = BidirectionalGRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_am)
# restore trained model
self.net.restore()
# rostopics
self.pub_reply = rospy.Publisher('reply', Reply, queue_size=10)
self.pub_complete = rospy.Publisher('complete_execute_scenario',
Empty,
queue_size=10)
rospy.Subscriber('raising_events', RaisingEvents,
self.handle_raise_events)
try:
rospy.wait_for_service('reception_db/query_data')
self.get_response_db = rospy.ServiceProxy(
'reception_db/query_data', DBQuery)
rospy.logwarn("waiting for reception DB module...")
except rospy.exceptions.ROSInterruptException as e:
rospy.logerr(e)
quit()
rospy.logwarn(
"network: {}, lang: {}, action_mask: {}, embedding: {}, user_number: {}"
.format(self.network_type, self.lang_type, self.is_am, self.is_emb,
self.user_num))
self.story.append('user number: %s' % self.user_num)
rospy.loginfo('\033[94m[%s]\033[0m initialized.' % rospy.get_name())
class Trainer():
def __init__(self, train_ratio=15/20, \
epochs=100, \
train_whole=True, \
):
self.train_ratio = train_ratio
self.epochs = epochs
#flag für fehleranalyse
self.startCountingErrors = False
self.count_all_one_eval = 0
self.count_all_one_train = 0
self.count_all_train = 0
self.count_all_eval = 0
#data.trainset = [(u,da_indx), ...]
#HIER acttemps und size
self.at = ActionTracker()
self.action_templates = self.at.action_templates
self.at.dialog_acts = self.at.dialog_acts
self.at.dialog_act_ids = self.at.dialog_act_ids
self.dataset, self.dialog_indices, self.other_da, self.other_da_id = self.at.get_data()
self.action_size = self.at.action_size
self.accs = []
self.loss = []
self.dev_accs =[]
self.dev_loss=[]
self.countErrors = Counter()
self.countPredicitions = 0
self.errorLog = {}
self.highest_accuracy = 0
#pickle.load(open("data/errors/errorLog.p", "rb"))
if train_whole:
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
#HIER numfeats
obs_size = self.emb.dim + self.bow_enc.vocab_size + self.action_size# + self.et.size_context_features
nb_hidden = 128
self.net = LSTM_net(obs_size,
self.action_size,
nb_hidden=nb_hidden)
def train(self):
print(self.other_da)
print('\n:: training started', \
'\ntrain len :\t', round(len(self.dialog_indices) * self.train_ratio), \
'\ndev len :\t', len(self.dialog_indices) - round(len(self.dialog_indices) * self.train_ratio))
print(self.epochs)
random.shuffle(self.dialog_indices)
self.dialog_indices_tr = self.dialog_indices[:round(len(self.dialog_indices) * self.train_ratio)]
self.dialog_indices_dev = self.dialog_indices[round(len(self.dialog_indices) * self.train_ratio):]
num_tr_examples = len(self.dialog_indices_tr)
for j in range(self.epochs):
#if j < 250:
# self.startCountingErrors == True
# iterate through dialogs
loss = 0.
random.shuffle(self.dialog_indices_tr)
for i, dialog_idx in enumerate(self.dialog_indices_tr):
# get start and end index
start, end = dialog_idx['start'], dialog_idx['end']
# train on dialogue
loss += self.dialog_train(self.dataset[start:end])
# print #iteration
sys.stdout.write('\r{}.[{}/{}]'.format(j+1, i+1, num_tr_examples))
print('\n\n:: {}.tr loss {}'.format(j+1, loss/num_tr_examples))
# evaluate every epoch
accuracy = self.evaluate()
self.accs.append(accuracy)
print(':: {}.dev accuracy {}\n'.format(j+1, accuracy))
#if accuracy > 0.98: # original threshold was 0.99
# self.net.save()
# plt.plot(self.accs)
# plt.xlabel("epochs")
# plt.ylabel('accuracy')
# plt.show()
# break
if j == self.epochs-1:
#print("ACC: {}".format(float(sum(self.accs[349:]))/float(len(self.accs[249:]))))
#self.countErrors["ACCURACY"]= float(sum(self.accs[349:]))/float(len(self.accs[249:]))
#self.errorLog[dt.datetime.now()] = self.countErrors
#if accuracy > 0.95:
self.net.save() # save any result after j iterations
#for key in self.countErrors.keys():
# print("ACC: {}".format(float(sum(self.accs[349:]))/float(len(self.accs[249:]))))
# self.countErrors[key] = self.countErrors[key]/ self.countPredicitions
#print(json.dumps(self.countErrors, indent=4))
#pickle.dump(self.accs, open("data/accuracies_trunc_09.p", "wb"))
#pickle.dump(self.loss, open("data/losses_trunc_09.p", "wb"))
#pickle.dump(self.at.dialog_acts, open("data/acts.p", "wb"))
#pickle.dump(self.at.dialog_act_ids, open("data/ids.p", "wb"))
#pickle.dump(self.dev_accs, open("data/dev_accuracies_00.p", "wb"))
#pickle.dump(self.dev_loss, open("data/dev_losses_00.p", "wb"))
print("Mean Accuracy: ", float(sum(self.accs[149:]))/float(len(self.accs[149:])))
print("Predictions LSTM Training: ", self.count_all_one_train)
print("LSTM ratio Training {}".format(float(self.count_all_one_train)/float(self.count_all_train)))
print("Predictions LSTM Evaluation: ", self.count_all_one_eval)
print("LSTM ratio Evaluation {}".format(float(self.count_all_one_eval)/float(self.count_all_eval)))
plt.plot(self.accs)
plt.xlabel("epochs")
plt.ylabel('loss')
plt.show()
break
def dialog_train(self, dialog):
self.at.reset_action_mask()
self.net.reset_state()
try:
self.at.set_exercise(self.at.get_dialog_act_id(dialog[0][1],action_template=True))
except:
print(dialog[0][1])
exit()
self.at.automaton.reset_current_state()
self.at.memory.reset_memory(self.at.exercise)
loss = 0.
self.at.initialize_am()
self.at.automaton.reset_current_state()
#error_code = re.compile(r"\[01\]+S")
for (u,r) in dialog:
if r.startswith("intro"):
r = "intro"
r = self.at.get_dialog_act_id(r)
#print(r)
u_emb = self.emb.encode(u)#HIER
#print('\nEmbedded Utterance, W2V:\n\n', u_emb)
u_bow = self.bow_enc.encode(u)#HIER
self.count_all_one_train += 1
action_mask = self.at.am
# forward propagation
# train step
features = np.concatenate((u_emb, u_bow, action_mask), axis=0)
lossValue , prediction = self.net.train_step(features, r, action_mask)
#current_state = self.at.automaton.current_state
#list_of_interest = [self.other_da[i] for i in range(len(self.other_da)) if action_mask[i] == 1]
#print("action mask: ",list_of_interest,"\ncurrent_state:",current_state.name,"\nprediction: ", self.other_da[prediction])
self.at.automaton.current_state,_ = self.at.walk(u, prediction)
#print(prediction)
#print(self.at.automaton.current_state.name)
self.count_all_train +=1
#print("predicted: ", self.at.dialog_acts[prediction])
loss += lossValue
#loss += self.net.train_step(features, r)
self.loss.append(loss/len(dialog))
return loss/len(dialog)
def evaluate(self):
self.net.reset_state()
dialog_accuracy = 0
dialog_loss = 0
errors = []
for dialog_idx in self.dialog_indices_dev:
start, end = dialog_idx['start'], dialog_idx['end']
dialog = self.dataset[start:end]
num_dev_examples = len(self.dialog_indices_dev)
# reset network
self.net.reset_state()
try:
self.at.set_exercise(self.at.get_dialog_act_id(dialog[0][1],action_template=True))
except:
print(dialog)
exit()
self.at.memory.reset_memory(self.at.exercise)
self.at.initialize_am()
# iterate through dialog
correct_examples = 0
#last_action = np.zeros([self.action_size], dtype=np.float32)
prev_prediction = -1
self.at.automaton.reset_current_state()
prev_state = ""
i = 0
loss = 0
for (u,r) in dialog:
if r.startswith("intro"):
r = "intro"
r = self.at.get_dialog_act_id(r)
i +=1
u_emb = self.emb.encode(u)#HIER
#print('\nEmbedded Utterance, W2V:\n\n', u_emb)
u_bow = self.bow_enc.encode(u)#HIER
prev_state = self.at.automaton.current_state.name
action_mask = self.at.am
features = np.concatenate((u_emb, u_bow, action_mask), axis=0)
prediction= self.net.forward(features, action_mask, r)
#current_state = self.at.automaton.current_state
#list_of_interest = [self.other_da[i] for i in range(len(self.other_da)) if action_mask[i] == 1]
#print("action mask: ",list_of_interest,"\ncurrent_state:",current_state.name,"\nprediction: ", self.other_da[prediction])
self.at.automaton.current_state,_ = self.at.walk(u,prediction)
#self.count_all_eval +=1
#if prediction != r:
# print("FROM: {}, with: {}, TO: {}, predicted: {}".format(prev_state, self.at.automaton.get_utterance_type(u), self.at.automaton.current_state.name, self.at.dialog_acts[prediction]))
#if self.startCountingErrors and prediction != r:
# self.countErrors[self.at.dialog_acts[prediction] + ":" + self.at.dialog_acts[r]] +=1
#if self.startCountingErrors:
# self.countPredicitions +=1
if prediction != r:
errors.append("EXERCISE: {}, step {}, Previous State: {}, Previous Act: {}, Prediction: {} Actual Da: {}, Utterance: {}, Symbol: {}".format(self.at.exercise, i, prev_state, self.at.dialog_acts[prev_prediction], self.at.dialog_acts[prediction], self.at.dialog_acts[r], u, self.at.automaton.get_utterance_type(u)))
prev_prediction = r
#print("Prediction: {} Actual Da: {}".format(self.at.dialog_acts[prediction],self.at.dialog_acts[r]))
#self.at.memory.update(prediction)
#last_action= np.zeros([self.action_size], dtype=np.float32)
#last_action[prediction-1] = 1
#prediction = self.net.forward(features)
#print("predicted: ", self.at.dialog_acts[prediction])
#if prediction == r:
# print(self.at.dialog_acts[r])
correct_examples += int(prediction == r)
self.count_all_eval += 1
#loss += loss_value[0]
# get dialog accuracy
dialog_accuracy += correct_examples/len(dialog)
#dialog_loss += loss / len(dialog)
if self.highest_accuracy > 0 and self.highest_accuracy -(dialog_accuracy/num_dev_examples) > 0.2:
for e in errors:
print(e)
print(self.highest_accuracy -(dialog_accuracy/num_dev_examples))
if self.highest_accuracy < dialog_accuracy/num_dev_examples:
self.highest_accuracy = dialog_accuracy/num_dev_examples
self.dev_loss.append(dialog_loss/num_dev_examples)
self.dev_accs.append(dialog_accuracy/num_dev_examples)
print("dev_loss: {}".format(dialog_loss/num_dev_examples))
return dialog_accuracy/num_dev_examples
class Trainer():
def __init__(self):
et = EntityTracker()
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
at = ActionTracker(et)
self.dataset, dialog_indices = Data(et, at).trainset
random.shuffle(dialog_indices)
self.dialog_indices_tr = dialog_indices[:200]
self.dialog_indices_dev = dialog_indices[200:250]
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
self.action_templates = at.get_action_templates()
action_size = at.action_size
nb_hidden = 128
self.net = LSTM_wrapper(obs_size=obs_size,
action_size=action_size,
nb_hidden=nb_hidden)
def train(self):
print('\n:: training started\n')
epochs = 80
for j in range(epochs):
# iterate through dialogs
num_tr_examples = len(self.dialog_indices_tr)
loss = 0.
for i,dialog_idx in enumerate(self.dialog_indices_tr):
# get start and end index
start, end = dialog_idx['start'], dialog_idx['end']
# train on dialogue
loss += self.dialog_train(self.dataset[start:end])
# print #iteration
sys.stdout.write('\r{}.[{}/{}]'.format(j+1, i+1, num_tr_examples))
print('\n\n:: {}.tr loss {}'.format(j+1, loss/num_tr_examples))
# evaluate every epoch
accuracy = self.evaluate()
print(':: {}.dev accuracy {}\n'.format(j+1, accuracy))
# if accuracy > 0.99:
# self.net.save()
# break
def dialog_train(self, dialog):
# create entity tracker
et = EntityTracker()
# create action tracker
at = ActionTracker(et)
# reset network
self.net.reset_state()
loss = 0.
# iterate through dialog
for (u,r) in dialog:
# print("Here in the dialog loop")
u_ent = et.extract_entities(u)
u_ent_features = et.context_features()
u_emb = self.emb.encode(u)
u_bow = self.bow_enc.encode(u)
# concat features
features = torch.autograd.Variable(torch.from_numpy(np.concatenate((u_ent_features, u_emb, u_bow), axis=0))).float()
# print(features)
# get action mask
action_mask = torch.autograd.Variable(torch.from_numpy(at.action_mask()))
r = torch.autograd.Variable(torch.LongTensor([r]))
# print(r)
# forward propagation
# train step
loss += self.net.train_step(features, r, action_mask)
return loss/len(dialog)
def evaluate(self):
# create entity tracker
et = EntityTracker()
# create action tracker
at = ActionTracker(et)
# reset network
self.net.reset_state()
dialog_accuracy = 0.
r_count = 0 #Count of task 15
count = 0 # Total count of rewards
for dialog_idx in self.dialog_indices_dev:
start, end = dialog_idx['start'], dialog_idx['end']
dialog = self.dataset[start:end]
num_dev_examples = len(self.dialog_indices_dev)
# create entity tracker
et = EntityTracker()
# create action tracker
at = ActionTracker(et)
# reset network
self.net.reset_state()
# iterate through dialog
correct_examples = 0
for (u,r) in dialog:
# encode utterance
u_ent = et.extract_entities(u)
u_ent_features = et.context_features()
u_emb = self.emb.encode(u)
u_bow = self.bow_enc.encode(u)
# concat features
# get action mask
features = torch.autograd.Variable(torch.from_numpy(np.concatenate((u_ent_features, u_emb, u_bow), axis=0))).float()
# print(features)
# get action mask
action_mask = torch.autograd.Variable(torch.from_numpy(at.action_mask()))
# r = torch.autograd.Variable(torch.LongTensor([r]))
# forward propagation
# train step
logits,probs,prediction = self.net.forward(features, action_mask)
# print("logits", logits)
# print("probs", probs)
# print("prediction", logits,probs,prediction)
# print(prediction,r)
correct_examples += int(prediction == r)
if r==15:
r_count += 1
count += 1
# get dialog accuracy
dialog_accuracy += correct_examples/len(dialog)
print("task 15 was the answer with freq",r_count/count)
return dialog_accuracy/num_dev_examples
def __init__(self, args):
self.response_accuracy = []
self.dialog_accuracy = []
try:
###################### selective import #############################
if args[0] == 'am':
self.is_action_mask = True
else:
self.is_action_mask = False
if args[1] == 'emb':
self.is_emb = True
else:
self.is_emb = False
self.network_type = args[2]
self.lang_type = args[3]
if self.lang_type == 'eng':
from modules.entities import EntityTracker
from modules.data_utils import Data
from modules.actions import ActionTracker
from modules.bow import BoW_encoder
elif self.lang_type == 'kor':
from modules.entities_kor import EntityTracker
from modules.data_utils_kor import Data
from modules.actions_kor import ActionTracker
from modules.bow_kor import BoW_encoder
###################################################################
except:
rospy.logwarn(
"please try again. i.e. ... train.py <am> <emb> <bidirectional_lstm> <eng>"
)
if self.is_emb:
if self.lang_type == 'eng':
self.emb = UtteranceEmbed(lang=self.lang_type)
elif self.lang_type == 'kor':
self.emb = UtteranceEmbed(lang=self.lang_type)
et = EntityTracker()
self.bow_enc = BoW_encoder()
at = ActionTracker(et)
at.do_display_template()
self.dataset, dialog_indices = Data(et, at).trainset
# self.dialog_indices_tr = dialog_indices[:200]
self.dialog_indices_tr = random.sample(dialog_indices, 200)
# self.dialog_indices_dev = dialog_indices[200:250]
self.dialog_indices_dev = random.sample(dialog_indices, 50)
self.action_templates = at.get_action_templates()
action_size = at.action_size
nb_hidden = 128
# set feature input
if self.is_action_mask and self.is_emb:
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features + at.action_size
elif self.is_action_mask and not (self.is_emb):
obs_size = self.bow_enc.vocab_size + et.num_features + at.action_size
elif not (self.is_action_mask) and self.is_emb:
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
elif not (self.is_action_mask) and not (self.is_emb):
obs_size = self.bow_enc.vocab_size + et.num_features
# set network_type type
if self.network_type == 'gru':
self.net = GRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
elif self.network_type == 'reversed_lstm':
self.net = ReversingLSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
elif self.network_type == 'reversed_gru':
self.net = ReversingGRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
elif self.network_type == 'stacked_gru':
self.net = StackedGRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
elif self.network_type == 'stacked_lstm':
self.net = StackedLSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
elif self.network_type == 'lstm':
self.net = LSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
elif self.network_type == 'bidirectional_lstm':
self.net = BidirectionalLSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
elif self.network_type == 'bidirectional_gru':
self.net = BidirectionalGRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
file_path = os.path.join(rospkg.RosPack().get_path('dialogue_system'),
'log', self.network_type)
# init logging
self.logger = self.get_logger(file_path)
msg = "'\033[94m[%s trainer]\033[0m initialized - %s (action_mask: %s, embedding: %s, lang: %s, obs_size: %s, bow: %s, context_feature: %s, action_size: %s)" % (
rospy.get_name(), self.net.__class__.__name__, self.is_action_mask,
self.is_emb, self.lang_type, obs_size, self.bow_enc.vocab_size,
et.num_features, action_size)
rospy.loginfo(msg)
class Trainer():
def __init__(self):
et = EntityTracker()
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
at = ActionTracker(et)
self.train_dataset, train_dialog_indices = Data(et, at).train_set
self.test_dataset, test_dialog_indices = Data(et, at).test_set
print('=========================\n')
print('length of Train dialog indices : ', len(train_dialog_indices))
print('=========================\n')
print('=========================\n')
print('length of Test dialog indices : ', len(test_dialog_indices))
print('=========================\n')
# Shuffle Training Dataset
random.shuffle(train_dialog_indices)
self.dialog_indices_tr = train_dialog_indices
self.dialog_indices_dev = test_dialog_indices
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
self.action_templates = at.get_action_templates()
action_size = at.action_size
# nb_hidden = 128
nb_hidden = 150
print('=========================\n')
print('Action_templates: ', action_size)
print('=========================\n')
self.net = LSTM_net(obs_size=obs_size,
action_size=action_size,
nb_hidden=nb_hidden)
self.et = et
self.at = at
action_projection = []
for action in self.action_templates:
action_projection.append(self.emb.encode(action))
self.action_projection = np.transpose(action_projection)
self.action_size = action_size
def train(self, exp_name, model_name):
print('\n:: training started\n')
epochs = 100
import joblib
per_response_list = []
per_dialogue_list = []
early_stop = False
early_stop_count = 0
for j in range(epochs):
# iterate through dialogs
num_tr_examples = len(self.dialog_indices_tr)
loss = 0.
for i,dialog_idx in enumerate(self.dialog_indices_tr):
# get start and end index
start, end = dialog_idx['start'], dialog_idx['end']
# train on dialogue
loss += self.dialog_train(self.train_dataset[start:end])
# print #iteration
sys.stdout.write('\r{}.[{}/{}]'.format(j+1, i+1, num_tr_examples))
print('\n\n:: {}.tr loss {}'.format(j+1, loss/num_tr_examples))
# evaluate every epoch
accuracy = self.evaluate()
per_response_list.append(accuracy[0])
per_dialogue_list.append(accuracy[1])
max_dialogue = max(per_dialogue_list)
if len(per_dialogue_list) > 1:
prev_max_dialogue = sorted(per_dialogue_list, reverse=True)[1]
if max_dialogue > 2.0 and accuracy[1] > prev_max_dialogue:
early_stop_count += 1
self.net.save(model_name)
print(':: {}.dev accuracy {}\n'.format(j+1, accuracy))
print('current max dialogue accuracy : {}\n'.format(sorted(per_dialogue_list, reverse=True)[0]))
print('Max Dialogue Accuracy : ', max(per_dialogue_list))
joblib.dump(per_response_list, 'emnlp_performance/with_slot/per_response_list_' + exp_name)
joblib.dump(per_dialogue_list, 'emnlp_performance/with_slot/per_dialogue_list_' + exp_name)
# self.net.save()
def dialog_train(self, dialog):
# create entity tracker
et = self.et
et.init_entities()
# create action tracker
at = self.at
# reset network
self.net.reset_state()
self.net.reset_attention()
loss = 0.
i = 0
pred_list = []
# iterate through dialog
for (u,r) in dialog:
i += 1
if r == '<UNK>':
continue
u_ent = et.extract_entities(u)
u_ent_features = et.context_features()
u_emb = self.emb.encode(u)
u_bow = self.bow_enc.encode(u)
# concat features
features = np.concatenate((u_ent_features, u_emb, u_bow), axis=0)
if i ==1:
loss += self.net.train_step(features, r, self.action_projection)
pred_list.append(r)
else:
action_one_hot = np.zeros(self.action_size)
action_one_hot[pred_list[-1]] = 1
loss += self.net.train_step(features, r, self.action_projection, action_one_hot)
pred_list.append(r)
return loss / len(dialog)
def evaluate(self):
dialog_accuracy = 0.
correct_dialogue_count = 0
# for each dialog
for dialog_idx in self.dialog_indices_dev:
start, end = dialog_idx['start'], dialog_idx['end']
dialog = self.test_dataset[start:end]
num_dev_examples = len(self.dialog_indices_dev)
# create entity tracker
et = self.et
et.init_entities()
# create action tracker
at = self.at
# reset network
self.net.reset_state()
self.net.reset_attention()
# iterate through dialog
correct_examples = 0
pred_list = []
i = 0
for (u,r) in dialog:
i += 1
if u == 'api_call no result':
correct_examples += 1
continue
if r == '<UNK>':
# correct_examples += 1
continue
# encode utterance
u_ent = et.extract_entities(u)
u_ent_features = et.context_features()
u_emb = self.emb.encode(u)
u_bow = self.bow_enc.encode(u)
# concat features
features = np.concatenate((u_ent_features, u_emb, u_bow), axis=0)
if i == 1:
prediction, user_attention_weights, action_weights = self.net.forward(features, self.action_projection)
pred_list.append(prediction)
else:
action_one_hot = np.zeros(self.action_size)
action_one_hot[pred_list[-1]] = 1
prediction, user_attention_weights, action_weights = self.net.forward(features, self.action_projection, action_one_hot)
pred_list.append(prediction)
correct_examples += int(prediction == r)
if correct_examples == len(dialog):
correct_dialogue_count += 1
# get dialog accuracy
dialog_accuracy += correct_examples / len(dialog)
per_response_accuracy = dialog_accuracy / num_dev_examples * 100
per_dialogue_accuracy = correct_dialogue_count / num_dev_examples * 100
print('=============================')
print('correct dialogue count')
print(correct_dialogue_count)
print('=============================\n')
return per_response_accuracy, per_dialogue_accuracy
class InteractiveSession:
def __init__(self):
self.et = EntityTracker()
self.at = ActionTracker(self.et)
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
obs_size = self.emb.dim + self.bow_enc.vocab_size + self.et.num_features
self.action_templates = self.at.get_action_templates()
action_size = self.at.action_size
nb_hidden = 128
self.net = LSTM_net(obs_size=obs_size,
action_size=action_size,
nb_hidden=nb_hidden)
# restore checkpoint
self.net.restore()
self.net.reset_state()
def reset(self):
self.net.reset_state()
self.et = EntityTracker()
self.at = ActionTracker(self.et)
def interact(self, utterance, intent, slot_values):
# get input from user
u = utterance.lower()
# check if user wants to begin new session
if u == 'clear' or u == 'reset' or u == 'restart':
self.reset()
return "reset successfully"
# check for entrance and exit command
elif u == 'exit' or u == 'stop' or u == 'quit' or u == 'q':
self.reset()
return "Thank you for using"
elif u == 'hello' or u == 'hi':
self.reset()
return "what can i do for you"
elif u == 'thank you' or u == 'thanks' or u == 'thank you very much':
self.reset()
return 'you are welcome'
else:
# encode
u_ent = self.et.extract_entities(u, intent, slot_values)
u_ent_features = self.et.context_features() # 5
u_emb = self.emb.encode(u) # 300
u_bow = self.bow_enc.encode(u) # 60
# concat features
features = np.concatenate((u_ent_features, u_emb, u_bow), axis=0)
# get action mask
action_mask = self.at.action_mask()
# action_mask = np.ones(self.net.action_size)
# forward
prediction = self.net.forward(features, action_mask)
response = self.action_templates[prediction]
if prediction == 0:
slot_values = copy.deepcopy(self.et.entities)
slot_values.pop('location')
memory = ', '.join(slot_values.values())
response = response.replace("memory", memory)
self.reset()
print('API CALL execute successfully and begin new session')
if prediction == 1:
response = response.replace("location",
self.et.entities['location'])
return response
class Trainer():
def __init__(self):
et = EntityTracker()
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
at = ActionTracker(et)
'''
['any preference on a type of cuisine', 'api_call <cuisine> <location> <party_size> <rest_type>', 'great let me do the reservation', 'hello what can i help you with today', 'here it is <info_address>', 'here it is <info_phone>', 'how many people would be in your party', "i'm on it", 'is there anything i can help you with', 'ok let me look into some options for you', 'sure is there anything else to update', 'sure let me find an other option for you', 'what do you think of this option: <restaurant>', 'where should it be', 'which price range are looking for', "you're welcome"]
'''
self.dataset, dialog_indices = Data(et, at).trainset
self.dialog_indices_tr = dialog_indices[:200]
self.dialog_indices_dev = dialog_indices[200:250]
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
self.action_templates = at.get_action_templates()
action_size = at.action_size
nb_hidden = 128
self.net = LSTM_net(obs_size=obs_size,
action_size=action_size,
nb_hidden=nb_hidden)
def train(self):
print('\n:: training started\n')
epochs = 20
for j in range(epochs):
# iterate through dialogs
num_tr_examples = len(self.dialog_indices_tr)
loss = 0.
for i, dialog_idx in enumerate(self.dialog_indices_tr):
# get start and end index
start, end = dialog_idx['start'], dialog_idx['end']
# train on dialogue
loss += self.dialog_train(self.dataset[start:end])
# print #iteration
sys.stdout.write('\r{}.[{}/{}]'.format(j + 1, i + 1,
num_tr_examples))
print('\n\n:: {}.tr loss {}'.format(j + 1, loss / num_tr_examples))
# evaluate every epoch
accuracy = self.evaluate()
print(':: {}.dev accuracy {}\n'.format(j + 1, accuracy))
if accuracy > 0.99:
self.net.save()
break
def dialog_train(self, dialog):
# create entity tracker
et = EntityTracker()
# create action tracker
at = ActionTracker(et)
# reset network
self.net.reset_state()
loss = 0.
# iterate through dialog
for (u, r) in dialog:
u_ent = et.extract_entities(u)
u_ent_features = et.context_features()
u_emb = self.emb.encode(u)
u_bow = self.bow_enc.encode(u)
# concat features
features = np.concatenate((u_ent_features, u_emb, u_bow), axis=0)
# get action mask
action_mask = at.action_mask()
# forward propagation
# train step
loss += self.net.train_step(features, r, action_mask)
return loss / len(dialog)
def evaluate(self):
# create entity tracker
et = EntityTracker()
# create action tracker
at = ActionTracker(et)
# reset network
self.net.reset_state()
dialog_accuracy = 0.
for dialog_idx in self.dialog_indices_dev:
start, end = dialog_idx['start'], dialog_idx['end']
dialog = self.dataset[start:end]
num_dev_examples = len(self.dialog_indices_dev)
# create entity tracker
et = EntityTracker()
# create action tracker
at = ActionTracker(et)
# reset network
self.net.reset_state()
# iterate through dialog
correct_examples = 0
for (u, r) in dialog:
# encode utterance
u_ent = et.extract_entities(u)
u_ent_features = et.context_features()
u_emb = self.emb.encode(u)
u_bow = self.bow_enc.encode(u)
# concat features
features = np.concatenate((u_ent_features, u_emb, u_bow),
axis=0)
# get action mask
action_mask = at.action_mask()
# forward propagation
# train step
prediction = self.net.forward(features, action_mask)
correct_examples += int(prediction == r)
# get dialog accuracy
dialog_accuracy += correct_examples / len(dialog)
return dialog_accuracy / num_dev_examples
class Trainer():
def __init__(self):
et = EntityTracker()
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
at = ActionTracker(et)
self.dataset, dialog_indices = Data(et, at).trainset
self.dialog_indices_tr = dialog_indices[:200]
self.dialog_indices_dev = dialog_indices[200:250]
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
self.action_templates = at.get_action_templates()
action_size = at.action_size
nb_hidden = 128
self.net = LSTM_net(obs_size=obs_size,
action_size=action_size,
nb_hidden=nb_hidden)
def train(self):
print('\n---training started---\n')
epochs = 20
for j in range(epochs):
# iterate through dialogs
num_tr_examples = len(self.dialog_indices_tr)
loss = 0.
for i,dialog_idx in enumerate(self.dialog_indices_tr):
# get start and end index
start, end = dialog_idx['start'], dialog_idx['end']
# train on dialogue
loss += self.dialog_train(self.dataset[start:end])
# print #iteration
sys.stdout.write('\r{}.[{}/{}]'.format(j+1, i+1, num_tr_examples))
print('\n\n--- {}.tr loss {} ---'.format(j+1, loss/num_tr_examples))
# evaluate every epoch
accuracy = self.evaluate()
print('--- {}.dev accuracy {} ---\n'.format(j+1, accuracy))
if accuracy > 0.99:
self.net.save()
break
def dialog_train(self, dialog):
# create entity tracker
et = EntityTracker()
# create action tracker
at = ActionTracker(et)
# reset network
self.net.reset_state()
loss = 0.
# iterate through dialog
for (u, r) in dialog:
u_ent = et.extract_entities(u)
u_ent_features = et.context_features()
u_emb = self.emb.encode(u)
u_bow = self.bow_enc.encode(u)
# concat features
features = np.concatenate((u_ent_features, u_emb, u_bow), axis=0)
# get action mask
action_mask = at.action_mask()
# forward propagation
# train step
loss += self.net.train_step(features, r, action_mask)
return loss/len(dialog)
def evaluate(self):
# create entity tracker
et = EntityTracker()
# create action tracker
at = ActionTracker(et)
# reset network
self.net.reset_state()
dialog_accuracy = 0.
for dialog_idx in self.dialog_indices_dev:
start, end = dialog_idx['start'], dialog_idx['end']
dialog = self.dataset[start:end]
num_dev_examples = len(self.dialog_indices_dev)
# create entity tracker
et = EntityTracker()
# create action tracker
at = ActionTracker(et)
# reset network
self.net.reset_state()
# iterate through dialog
correct_examples = 0
for (u,r) in dialog:
# encode utterance
u_ent = et.extract_entities(u)
u_ent_features = et.context_features()
u_emb = self.emb.encode(u)
u_bow = self.bow_enc.encode(u)
# concat features
features = np.concatenate((u_ent_features, u_emb, u_bow), axis=0)
# get action mask
action_mask = at.action_mask()
# forward propagation
# train step
prediction = self.net.forward(features, action_mask)
correct_examples += int(prediction == r)
# get dialog accuracy
dialog_accuracy += correct_examples/len(dialog)
return dialog_accuracy/num_dev_examples
class Train():
def __init__(self, args):
self.response_accuracy = []
self.dialog_accuracy = []
try:
###################### selective import #############################
if args[0] == 'am':
self.is_action_mask = True
else:
self.is_action_mask = False
if args[1] == 'emb':
self.is_emb = True
else:
self.is_emb = False
self.network_type = args[2]
self.lang_type = args[3]
if self.lang_type == 'eng':
from modules.entities import EntityTracker
from modules.data_utils import Data
from modules.actions import ActionTracker
from modules.bow import BoW_encoder
elif self.lang_type == 'kor':
from modules.entities_kor import EntityTracker
from modules.data_utils_kor import Data
from modules.actions_kor import ActionTracker
from modules.bow_kor import BoW_encoder
###################################################################
except:
rospy.logwarn(
"please try again. i.e. ... train.py <am> <emb> <bidirectional_lstm> <eng>"
)
if self.is_emb:
if self.lang_type == 'eng':
self.emb = UtteranceEmbed(lang=self.lang_type)
elif self.lang_type == 'kor':
self.emb = UtteranceEmbed(lang=self.lang_type)
et = EntityTracker()
self.bow_enc = BoW_encoder()
at = ActionTracker(et)
at.do_display_template()
self.dataset, dialog_indices = Data(et, at).trainset
# self.dialog_indices_tr = dialog_indices[:200]
self.dialog_indices_tr = random.sample(dialog_indices, 200)
# self.dialog_indices_dev = dialog_indices[200:250]
self.dialog_indices_dev = random.sample(dialog_indices, 50)
self.action_templates = at.get_action_templates()
action_size = at.action_size
nb_hidden = 128
# set feature input
if self.is_action_mask and self.is_emb:
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features + at.action_size
elif self.is_action_mask and not (self.is_emb):
obs_size = self.bow_enc.vocab_size + et.num_features + at.action_size
elif not (self.is_action_mask) and self.is_emb:
obs_size = self.emb.dim + self.bow_enc.vocab_size + et.num_features
elif not (self.is_action_mask) and not (self.is_emb):
obs_size = self.bow_enc.vocab_size + et.num_features
# set network_type type
if self.network_type == 'gru':
self.net = GRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
elif self.network_type == 'reversed_lstm':
self.net = ReversingLSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
elif self.network_type == 'reversed_gru':
self.net = ReversingGRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
elif self.network_type == 'stacked_gru':
self.net = StackedGRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
elif self.network_type == 'stacked_lstm':
self.net = StackedLSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
elif self.network_type == 'lstm':
self.net = LSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
elif self.network_type == 'bidirectional_lstm':
self.net = BidirectionalLSTM(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
elif self.network_type == 'bidirectional_gru':
self.net = BidirectionalGRU(obs_size=obs_size,
nb_hidden=nb_hidden,
action_size=action_size,
lang=self.lang_type,
is_action_mask=self.is_action_mask)
file_path = os.path.join(rospkg.RosPack().get_path('dialogue_system'),
'log', self.network_type)
# init logging
self.logger = self.get_logger(file_path)
msg = "'\033[94m[%s trainer]\033[0m initialized - %s (action_mask: %s, embedding: %s, lang: %s, obs_size: %s, bow: %s, context_feature: %s, action_size: %s)" % (
rospy.get_name(), self.net.__class__.__name__, self.is_action_mask,
self.is_emb, self.lang_type, obs_size, self.bow_enc.vocab_size,
et.num_features, action_size)
rospy.loginfo(msg)
def train(self, cont=False):
# logging and print
msg = "training started."
rospy.loginfo(msg)
# call previous trained model
if cont:
self.net.restore()
epochs = 20
# start measuring time
for j in range(epochs):
num_tr_examples = len(self.dialog_indices_tr)
loss = 0.
for i, dialog_idx in enumerate(self.dialog_indices_tr):
start, end = dialog_idx['start'], dialog_idx['end']
# train dialog
loss += self.dialog_train(self.dataset[start:end])
sys.stdout.write('\r{}.[{}/{}]'.format(j + 1, i + 1,
num_tr_examples))
# logging and print
msg = '\n\n {}.tr loss {}'.format(j + 1, loss / num_tr_examples)
rospy.loginfo(msg)
turn_accuracy, dialog_accuracy = self.evaluate()
msg = '\n{}.dev turn_accuracy {}, dialog_accuracy {}'.format(
j + 1, turn_accuracy, dialog_accuracy)
rospy.loginfo(msg)
if dialog_accuracy > 0.999:
self.net.save()
break
# save checkpoint
self.net.save()
def dialog_train(self, dialog):
###################################################################
if self.lang_type == 'eng':
from modules.entities import EntityTracker
from modules.data_utils import Data
from modules.actions import ActionTracker
from modules.bow import BoW_encoder
elif self.lang_type == 'kor':
from modules.entities_kor import EntityTracker
from modules.data_utils_kor import Data
from modules.actions_kor import ActionTracker
from modules.bow_kor import BoW_encoder
###################################################################
et = EntityTracker()
at = ActionTracker(et)
# reset state in network_type
self.net.reset_state()
loss = 0.
for (u, r) in dialog:
u_ent = et.extract_entities(u)
u_ent_features = et.context_features()
u_bow = self.bow_enc.encode(u)
if self.is_emb:
u_emb = self.emb.encode(u)
if self.is_action_mask:
action_mask = at.action_mask()
# print(u, r)
# print(u_ent_features)
# print('================================')
# print(u_emb)
# print('================================')
# print(u_bow)
# print('================================')
# print(action_mask)
# concatenated features
if self.is_action_mask and self.is_emb:
features = np.concatenate(
(u_ent_features, u_emb, u_bow, action_mask), axis=0)
elif self.is_action_mask and not (self.is_emb):
features = np.concatenate((u_ent_features, u_bow, action_mask),
axis=0)
elif not (self.is_action_mask) and self.is_emb:
features = np.concatenate((u_ent_features, u_emb, u_bow),
axis=0)
elif not (self.is_action_mask) and not (self.is_emb):
features = np.concatenate((u_ent_features, u_bow), axis=0)
# forward propagation with cumulative loss
if self.is_action_mask:
loss += self.net.train_step(features, r, action_mask)
else:
loss += self.net.train_step(features, r)
return loss / len(dialog)
def evaluate(self, eval=False):
###################################################################
if self.lang_type == 'eng':
from modules.entities import EntityTracker
from modules.data_utils import Data
from modules.actions import ActionTracker
from modules.bow import BoW_encoder
elif self.lang_type == 'kor':
from modules.entities_kor import EntityTracker
from modules.data_utils_kor import Data
from modules.actions_kor import ActionTracker
from modules.bow_kor import BoW_encoder
###################################################################
et = EntityTracker()
at = ActionTracker(et)
# only for evaluation purpose
if eval:
self.net.restore()
# reset entities extractor
turn_accuracy = 0.
dialog_accuracy = 0.
for dialog_idx in self.dialog_indices_dev:
start, end = dialog_idx['start'], dialog_idx['end']
dialog = self.dataset[start:end]
num_dev_examples = len(self.dialog_indices_dev)
et = EntityTracker()
at = ActionTracker(et)
# reset network_type before evaluate.
self.net.reset_state()
correct_examples = 0
for (u, r) in dialog:
u_ent = et.extract_entities(u)
u_ent_features = et.context_features()
u_bow = self.bow_enc.encode(u)
if self.is_emb:
u_emb = self.emb.encode(u)
if self.is_action_mask:
action_mask = at.action_mask()
# concatenated features
if self.is_action_mask and self.is_emb:
features = np.concatenate(
(u_ent_features, u_emb, u_bow, action_mask), axis=0)
elif self.is_action_mask and not (self.is_emb):
features = np.concatenate(
(u_ent_features, u_bow, action_mask), axis=0)
elif not (self.is_action_mask) and self.is_emb:
features = np.concatenate((u_ent_features, u_emb, u_bow),
axis=0)
elif not (self.is_action_mask) and not (self.is_emb):
features = np.concatenate((u_ent_features, u_bow), axis=0)
if self.is_action_mask:
probs, prediction = self.net.forward(features, action_mask)
else:
probs, prediction = self.net.forward(features)
correct_examples += int(prediction == r)
turn_accuracy += correct_examples / len(dialog)
accuracy = correct_examples / len(dialog)
if (accuracy == 1.0):
dialog_accuracy += 1
turn_accuracy = turn_accuracy / num_dev_examples
dialog_accuracy = dialog_accuracy / num_dev_examples
return turn_accuracy, dialog_accuracy
def get_logger(self, filename):
"""Return a logger instance that writes in filename
Args:
filename: (string) path to log.txt
Returns:
logger: (instance of logger)
"""
logger = logging.getLogger('logger')
logger.setLevel(logging.DEBUG)
logging.basicConfig(format='%(message)s', level=logging.DEBUG)
handler = logging.FileHandler(filename)
handler.setLevel(logging.DEBUG)
handler.setFormatter(
logging.Formatter('%(asctime)s:%(levelname)s: %(message)s'))
logging.getLogger().addHandler(handler)
return logger
class InteractiveSession():
def __init__(self):
self.bow_enc = BoW_encoder()
self.emb = UtteranceEmbed()
self.at = ActionTracker(interact=True)
print(self.at.dialog_acts)
_, _, self.other_da, self.other_da_id = self.at.get_data()
self.action_templates = self.at.action_templates
self.action_size = self.at.action_size
#self.responses = [r[0] for r in self.at.responses]
#print(self.emb.dim , self.bow_enc.vocab_size , self.action_size)
self.obs_size = self.emb.dim + self.bow_enc.vocab_size + self.action_size
nb_hidden = 128
self.exercises = list(self.at.action_templates.keys())
self.net = LSTM_net(obs_size=self.obs_size,
action_size=self.action_size,
nb_hidden=nb_hidden)
# restore checkpoint
self.net.restore()
# setup remote interaction via zeromq
context = zmq.Context.instance()
# PUB socket for sending system-output
self.pubSocket: zmq.Socket = context.socket(zmq.PUB)
self.pubSocket.bind(Config.SYSTEM_OUTPUT_ADDR)
# SUB socket for receiving user-input
self.subSocket: zmq.Socket = context.socket(zmq.SUB)
self.subSocket.connect(Config.USER_INPUT_ADDR)
self.subSocket.setsockopt_string(zmq.SUBSCRIBE,
Config.USER_INPUT_SUBJECT)
self.lastOutputMessage: str = None
self.requestHandler: threading.Thread = None
self.requestInterrupted: threading.Lock = None
def has_exercise(self, exercise_name: str) -> bool:
return exercise_name in self.at.action_templates
def get_current_exercise(self) -> str:
return self.at.exercise
def send(self, message, utteranceType, state, dialogAct):
print(">>", message)
data = SystemOutputData(text=message,
utterance_type=utteranceType,
state=state,
dialogue_act=dialogAct)
msg = SystemOutputMessage(type=MessageType.OUTPUT,
timestamp=int(time.time() * 1000),
data=data)
jsonMsg = json.dumps(msg.to_dict())
self.lastOutputMessage = jsonMsg
rawMsg = "%s %s" % (to_enum(MessageType, msg.type), jsonMsg)
# print(" ", rawMsg)#DEBUG
self.pubSocket.send_string(rawMsg)
def receive(self) -> 'UserInputMessage':
rawMsg = self.subSocket.recv_string()
# print(" ", rawMsg)#DEBUG
# FIXME detect & handle when message has other type than "input"
# TODO handle message types other than "input"
msgTypeStr = to_enum(MessageType, MessageType.INPUT)
# extract JSON message by removing subject/subscription string from beginning of message:
msgStr = rawMsg.replace(msgTypeStr, "", 1).strip()
return user_input_message_from_dict(json.loads(msgStr))
def start_processing_requests(self, verbose: bool = False):
self.stop_processing_requests()
self.requestInterrupted = threading.Lock()
self.requestInterrupted.acquire()
self.requestHandler = threading.Thread(
target=request_worker,
args=(Config.QUERY_SYSTEM_OUTPUT_ADDR, self,
self.requestInterrupted, verbose))
self.requestHandler.start()
def stop_processing_requests(self):
if self.requestInterrupted is not None and self.requestHandler is not None and self.requestHandler.isAlive(
):
self.requestInterrupted.release()
# print("should stop request_worker thread now!")
self.requestHandler.join()
print("stopped request_worker thread.")
def interact_anew(self,
verbose,
update,
out_file,
new_state_name: str = None):
prev_prediction = -1
self.net.reset_state()
if new_state_name is None:
new_state_name = random.choice(self.exercises)
self.at.set_exercise_string(new_state_name)
self.at.memory.reset_memory(self.at.exercise)
self.at.initialize_am()
u = "start"
if verbose == True:
print("\n*******************\nÄUßERUNGSTYP: ",
self.at.automaton.get_utterance_type(u))
action_mask = self.at.am
u_emb = self.emb.encode(u) # HIER
u_bow = self.bow_enc.encode(u) # HIER
features = np.concatenate((u_emb, u_bow, action_mask), axis=0)
# forward
prediction = self.net.forward(features,
action_mask,
self.other_da,
print_top=True)
self.at.automaton.current_state, response = self.at.walk(u, prediction)
prev_prediction = prediction
if verbose == True:
print("AKTUELLER ZUSTAND: ", self.at.automaton.current_state.name,
"\nDIALOGAKT: ", self.other_da[prediction],
"\n*******************\n"
) # print("CURRENT Exercise: ", self.at.exercise)
# self.net.train_step(features, prediction, action_mask)
# response = self.at.get_text(prediction)
self.send(response, self.at.automaton.get_utterance_type(u),
self.at.automaton.current_state.name,
self.at.dialog_acts[prediction]) # print('>>', response)
if out_file:
out_file.write(u + '\t' + response + '\t' +
self.other_da[prediction] + '\n')
return prev_prediction
def interact(self, verbose, update, write_to_file):
print(
"How do you do? I'm glad you're here. Let's do some exercises together."
)
# open file to write around here
# only if write_to_file !
with open('data/dialogs_from_interaction', 'a') as out:
with open("data/utterances_from_interaction", "a") as uttout:
self.start_processing_requests(verbose)
out_file = None
if write_to_file:
out_file = out
prev_prediction = self.interact_anew(verbose, update, out_file)
# begin interaction loop
while True:
if verbose == True:
print([
self.other_da[i] for i in range(len(self.at.am))
if self.at.am[i] == 1
])
#print("prediction: ",self.at.dialog_acts[prediction])
#print("response: ",response)
# get input from user; utterance isn't saved!
message = self.receive() # u = input(':: ')
u = message.data.strip()
print(":: %s" % (u))
# check if user wants to begin new session
if u == 'clear' or u == 'reset' or u == 'restart' or u == 'von vorne' or u == 'neustart':
if write_to_file:
out.write("WARNING: Interaction was reset.\n\n")
prev_prediction = self.interact_anew(
verbose, update, out_file)
if write_to_file:
uttout.write(
u + "\t" +
self.at.automaton.get_utterance_type(u) + "\n")
elif self.other_da[prev_prediction] in [
"closing", "finish"
]:
if write_to_file:
out.write('\n')
prev_prediction = self.interact_anew(
verbose, update, out_file)
# check for exit command
elif u == 'exit' or u == 'quit':
if update == True:
self.net.save()
self.send(
'Ok, wir brechen ab. Bis zum nächsten Mal!',
'EXIT COMMAND',
self.at.automaton.current_state.name, "UNKNOWN"
) # print('>> Ok, wir brechen ab. Bis zum nächsten Mal!')
if write_to_file:
out.write("WARNING: Dialog may not be finished.\n")
break
else:
if u == "\n" or u == "":
u = '<SILENCE>'
#self.at.reset_action_mask()
if self.at.automaton.current_state.name == "all_one":
if verbose == True:
print("\n*******************\nÄUßERUNGSTYP: ",
self.at.automaton.get_utterance_type(u))
##print(self.at.automaton.current_state.name, self.at.dialog_acts[prev_prediction])
else:
if verbose == True:
print("\n*******************\nÄUßERUNGSTYP: ",
self.at.automaton.get_utterance_type(u))
action_mask = self.at.am
u_emb = self.emb.encode(u) #HIER
u_bow = self.bow_enc.encode(u) #HIER
features = np.concatenate((u_emb, u_bow, action_mask),
axis=0)
# forward
prediction = self.net.forward(features,
action_mask,
self.other_da,
print_top=True)
self.at.automaton.current_state, response = self.at.walk(
u, prediction)
if prediction != "all_one":
prev_prediction = prediction
#self.at.memory.update(prediction)
if verbose == True:
print(
"AKTUELLER ZUSTAND: ",
self.at.automaton.current_state.name,
"\nDIALOGAKT: ", self.other_da[prediction],
"\n*******************\n"
) #print("CURRENT Exercise: ", self.at.exercise)
self.net.train_step(features, prediction, action_mask)
self.send(response,
self.at.automaton.get_utterance_type(u),
self.at.automaton.current_state.name,
self.at.dialog_acts[prediction]
) # print('>>', response)
if write_to_file:
out.write(u + '\t' + response + '\t' +
self.other_da[prediction] + '\n')
uttout.write(
u + "\t" +
self.at.automaton.get_utterance_type(u) + "\n")
if write_to_file:
out.write('\n')
self.stop_processing_requests()
