class DIALOG(nn.Module):
def __init__(self, vocab, db, config):
super(DIALOG, self).__init__()
self.vocab = vocab
self.db = db
self.pointer_size = config.pointer_size
self.max_belief_len = config.max_belief_len
self.nlu = BERTNLU() # fixed, not finetuning
self.context_encoder = ContextEncoder(vocab.vocab_size, config.hidden_size, config.hidden_size, config.dropout, config.num_layers, vocab.word2idx["<pad>"])
self.belief_decoder = BeliefDecoder(vocab, self.context_encoder.embedding, config.hidden_size, config.dropout, config.num_layers, config.max_value_len)
self.policy = RulePolicy()
self.nlg = TemplateNLG(is_user=False)
# self.action_decoder = ActionDecoder(vocab, self.context_encoder.embedding, config.hidden_size, config.pointer_size, config.dropout, config.num_layers, config.max_act_len)
# self.response_decoder = ResponseDecoder(vocab, self.context_encoder.embedding, config.hidden_size, config.pointer_size, config.dropout, config.num_layers, config.max_sentence_len)
self.load_embedding() # load Glove & Kazuma embedding
def forward(self, turn_inputs, turn_contexts, turn_context_lengths, action_history, teacher_forcing, test=False):
"""
turn_inputs: {
"user": [batch, len]
"response": [batch, len]
"belief": [batch, slots, len]
"gate": [batch, slots]
"action": [batch, len]
"usr": [batch] => string list
"context: [batch, turns, 2] => string list
"prev_gate": [batch, slots]
}
turn_contexts: [batch, len]
turn_context_lengths: [batch]
action_history: [batch, actions, 4] => string list
"""
if test:
return self.test_forward()
else:
return self.train_forward(turn_inputs, turn_contexts, turn_context_lengths, action_history, teacher_forcing)
def train_forward(self, turn_inputs, turn_contexts, turn_context_lengths, action_history, teacher_forcing):
batch_size = turn_contexts.size(0)
user = turn_inputs["usr"]
history = turn_inputs["context"]
for batch_idx in range(batch_size):
self.states[batch_idx]["history"].append(["usr", turn_inputs["usr"][batch_idx]])
input_action = self.nlu.predict(user[batch_idx], history[batch_idx]) # [[intent, domain, slot, value], ..., [intent, domain, slot, value]]
self.states[batch_idx]["user_action"] = input_action
action_history[batch_idx] += input_action
# update request state
for action in input_action:
intent, domain, slot, _ = action
slot, _ = clean_slot_values(domain, slot, _)
domain = domain.lower()
if intent == "Request":
if not self.states[batch_idx]["request_state"].get(domain):
self.states[batch_idx]["request_state"][domain] = {}
self.states[batch_idx]["request_state"][domain][slot] = 0
# calculate score of input action generated by BERT NLU
encoded_action_history = []
history_lengths = []
for batch_idx, actions in enumerate(action_history):
encoded_actions = []
for action in actions:
intent, domain, slot, value = action
slot = ontology.normlize_slot_names.get(slot, slot)
encoded_actions += self.vocab.encode(" ".join([intent, domain, slot, value]))[1:-1]
encoded_actions = [self.vocab.word2idx["<bos>"]] + encoded_actions + [self.vocab.word2idx["<eos>"]]
history_lengths.append(len(encoded_actions))
encoded_action_history.append(encoded_actions)
history_lengths = torch.tensor(history_lengths).cuda()
encoded_action_history_ = torch.zeros((batch_size, history_lengths.max().item()), dtype=torch.int64).cuda()
for batch_idx, actions in enumerate(encoded_action_history):
encoded_action_history_[batch_idx, :history_lengths[batch_idx]] = torch.tensor(actions)
action_history_outputs, action_history_hidden = self.context_encoder(encoded_action_history_, history_lengths) # [batch, len, hidden], [layers, batch, hidden]
encoder_outputs_, encoder_hidden_ = self.context_encoder(turn_contexts, turn_context_lengths) # [batch, len, hidden], [layers, batch, hidden]
action_score = torch.sigmoid(torch.matmul(action_history_hidden[-1].unsqueeze(dim=1), encoder_hidden_[-1].unsqueeze(dim=2))) # [batch, 1, 1]
# action score for attention score
action_score_attenion = torch.zeros(batch_size, (turn_context_lengths+history_lengths).max().item()).cuda() # [batch, len]
for batch_idx, context_len in enumerate(turn_context_lengths):
action_score_attenion[batch_idx, :context_len] = 1-action_score.squeeze()[batch_idx]
action_score_attenion[batch_idx, context_len:context_len+history_lengths[batch_idx].item()] = action_score.squeeze()[batch_idx]
# weighted sum & weighted concat
encoder_outputs, turn_contexts, turn_context_lengths = self.concat(encoder_outputs_, action_history_outputs, turn_contexts, turn_context_lengths, \
encoded_action_history_, history_lengths)
action_score = action_score.transpose(0,1).contiguous()
encoder_hidden = (1 - action_score) * encoder_hidden_ + action_score * action_history_hidden # [layers, batch, hidden]
gate_outputs, all_probs, all_pred_words = self.belief_decoder(encoder_outputs, encoder_hidden, turn_contexts, turn_context_lengths, \
turn_inputs["belief"], teacher_forcing, action_score_attenion) # [batch, slots, 3], [batch, slots, len, vocab], [batch, slots, len], [batch, slots]
gate_preds = gate_outputs.argmax(dim=2)
# prev_gate = turn_inputs.get("prev_gate")
max_value_len = 0
belief_gen = [] # [batch, slots, len]
for batch_idx, batch in enumerate(all_pred_words):
belief_gen_ = [] # [slots, len]
for slot_idx, pred_words in enumerate(batch):
if gate_preds[batch_idx, slot_idx].item() == ontology.gate_dict["none"]:
belief_gen_.append(self.vocab.encode("none")[1:])
len_ = len(self.vocab.encode("none")[1:])
elif gate_preds[batch_idx, slot_idx].item() == ontology.gate_dict["don't care"]:
belief_gen_.append(self.vocab.encode("don't care")[1:])
len_ = len(self.vocab.encode("don't care")[1:])
else:
for idx, value in enumerate(pred_words):
if value == self.vocab.word2idx["<eos>"]:
break
belief_gen_.append(pred_words[:idx+1].tolist())
len_ = idx + 1
max_value_len = max(max_value_len, len_)
belief_gen.append(belief_gen_)
gate_label = turn_inputs["gate"]
gate_loss = F.cross_entropy(gate_outputs.view(-1, 3), gate_label.view(-1))
# if prev_gate is not None:
# turn_domain = self.make_turn_domain(prev_gate, gate_preds) # [batch]
# else:
# turn_domain = self.make_turn_domain(None, gate_preds)
# prev_gate = gate_preds.detach() # [batch, slots]
acc_belief = torch.ones(batch_size, len(ontology.all_info_slots)).cuda()
gate_mask = (gate_label != gate_preds)
acc_belief.masked_fill_(gate_mask, value=0) # fail to predict gate
value_label = turn_inputs["belief"]
value_label_lengths = torch.zeros(batch_size, len(ontology.all_info_slots), dtype=torch.int64).cuda()
for batch_idx, batch in enumerate(value_label):
for slot_idx, pred_words in enumerate(batch):
value = pred_words[pred_words != self.vocab.word2idx["<pad>"]].tolist() # remove padding
value_label_lengths[batch_idx, slot_idx] = len(value)
if value != belief_gen[batch_idx][slot_idx]:
acc_belief[batch_idx, slot_idx] = 0 # fail to predict value
if teacher_forcing:
value_loss = masked_cross_entropy_for_value(all_probs, value_label, value_label_lengths)
else:
min_len = min(value_label.size(2), all_probs.size(2))
value_loss = masked_cross_entropy_for_value(all_probs[:, :, :min_len, :].contiguous(), value_label[:, :, :min_len].contiguous(), value_label_lengths)
### make state for Rule policy
# for batch_idx, belief in enumerate(belief_gen):
# for slot_idx, slot in enumerate(belief):
# value = self.vocab.decode(slot[:-1])
# domain, slot = ontology.all_info_slots[slot_idx].split("-")
# # book slots
# for slot_ in self.states[batch_idx]["belief_state"][domain]["book"].keys():
# if slot_ == "booked":
# continue
# slot_ = ontology.normlize_slot_names.get(slot_, slot_)
# if slot_ == slot:
# self.states[batch_idx]["belief_state"][domain]["book"][slot_] = value
# # semi slots
# for slot_ in self.states[batch_idx]["belief_state"][domain]["semi"].keys():
# slot_ = ontology.normlize_slot_names.get(slot_, slot_)
# if slot_ == slot:
# self.states[batch_idx]["belief_state"][domain]["semi"][slot_] = value
# ### policy
# output_actions = []
# for batch_idx, state in enumerate(self.states):
# output_actions.append(self.policy.predict(state))
# ### NLG
# model_responses = []
# for batch_idx, output_action in enumerate(output_actions):
# model_responses.append(self.nlg.generate(output_actions[batch_idx]))
return gate_loss, value_loss, acc_belief, belief_gen, action_history
def load_embedding(self):
glove = GloveEmbedding()
kazuma = KazumaCharEmbedding()
embed = self.context_encoder.embedding.weight.data
for word, idx in self.vocab.word2idx.items():
embed[idx] = torch.tensor(glove.emb(word, default="zero") + kazuma.emb(word, default="zero"))
# self.context_encoder.embedding.weight.data = embed
# self.belief_decoder.slot_embedding.weight.data = embed
def make_turn_domain(self, prev_gate, gate):
batch_size = gate.size(0)
turn_domain = torch.zeros((batch_size, len(ontology.all_domains))).cuda()
if prev_gate is None: # first turn
turn_gate = (gate != ontology.gate_dict["none"]).long()
else:
turn_gate = (gate != prev_gate).long() # find changed gate
for slot_idx in range(len(ontology.all_info_slots)):
domain, slot = ontology.all_info_slots[slot_idx].split("-")
domain_idx = ontology.all_domains.index(domain)
turn_domain[:, domain_idx] += turn_gate[:, slot_idx]
turn_domain = turn_domain.argmax(dim=1).tolist() # [batch]
return turn_domain
def parse_action(self, action):
"""
action: [len] => list
"""
domains = ontology.domain_action_slot["domain"]
actions = ontology.domain_action_slot["action"]
slots = ontology.domain_action_slot["slot"]
parsed_action = []
decoded_action = []
for token in action:
decoded_action.append(self.vocab.idx2word[token])
for token in decoded_action:
if token in domains:
domain = token
if token in actions:
action = token
if domain == "general":
parsed_action.append("{}-{}".format(domain, action))
if token in slots:
slot = token
parsed_action.append("{}-{}-{}".format(domain, action, slot))
return parsed_action
def concat(self, encoder_outputs, action_history_outputs, contexts, context_lengths, action_history, history_lengths):
batch_size = contexts.size(0)
hidden_size = encoder_outputs.size(2)
lengths = context_lengths + history_lengths
new_contexts = torch.zeros(size=(batch_size, lengths.max().item()), dtype=torch.int64).cuda()
new_encoder_outputs = torch.zeros(size=(batch_size, lengths.max().item(), hidden_size)).cuda()
for batch_idx in range(batch_size):
new_contexts[batch_idx, :lengths[batch_idx]] = torch.cat([contexts[batch_idx, :context_lengths[batch_idx]], \
action_history[batch_idx, :history_lengths[batch_idx]]], dim=0) # [batch, len]
new_encoder_outputs[batch_idx, :lengths[batch_idx], :] = torch.cat([encoder_outputs[batch_idx, :context_lengths[batch_idx], :], \
action_history_outputs[batch_idx, :history_lengths[batch_idx], :]], dim=0) # [batch, len, hidden]
return new_encoder_outputs, new_contexts, lengths
class Dialog(Agent):
def __init__(self, model_file=DEFAULT_MODEL_URL, name="Dialog"):
super(Dialog, self).__init__(name=name)
if not os.path.exists(os.path.join(DEFAULT_DIRECTORY,'multiwoz/data')):
os.mkdir(os.path.join(DEFAULT_DIRECTORY,'multiwoz/data'))
### download multiwoz data
print('down load data from', DEFAULT_ARCHIVE_FILE_URL)
if not os.path.exists(os.path.join(DEFAULT_DIRECTORY,'multiwoz/save')):
os.mkdir(os.path.join(DEFAULT_DIRECTORY,'multiwoz/save'))
### download trained model
print('down load model from', DEFAULT_MODEL_URL)
model_path = ""
config = Config()
parser = config.parser
config = parser.parse_args()
with open("assets/never_split.txt") as f:
never_split = f.read().split("\n")
self.tokenizer = BertTokenizer("assets/vocab.txt", never_split=never_split)
self.nlu = BERTNLU()
self.dst_ = DST(config).cuda()
ckpt = torch.load("save/model_Sun_Jun_21_07:08:48_2020.pt", map_location = lambda storage, loc: storage.cuda(local_rank))
self.dst_.load_state_dict(ckpt["model"])
self.dst_.eval()
self.policy = RulePolicy()
self.nlg = TemplateNLG(is_user=False)
self.init_session()
self.slot_mapping = {
"leave": "leaveAt",
"arrive": "arriveBy"
}
def init_session(self):
self.nlu.init_session()
self.policy.init_session()
self.nlg.init_session()
self.history = []
self.state = default_state()
pass
def response(self, user):
self.history.append(["user", user])
user_action = []
self.input_action = self.nlu.predict(user, context=[x[1] for x in self.history[:-1]])
self.input_action = deepcopy(self.input_action)
for act in self.input_action:
intent, domain, slot, value = act
if intent == "Request":
user_action.append(act)
if not self.state["request_state"].get(domain):
self.state["request_state"][domain] = {}
if slot not in self.state["request_state"][domain]:
self.state['request_state'][domain][slot] = 0
context = " ".join([utterance[1] for utterance in self.history])
context = context[-MAX_CONTEXT_LENGTH:]
context = self.tokenizer.encode(context)
context = torch.tensor(context, dtype=torch.int64).unsqueeze(dim=0).cuda() # [1, len]
belief_gen = self.dst_(None, context, 0, test=True)[0] # [slots, len]
for slot_idx, domain_slot in enumerate(ontology.all_info_slots):
domain, slot = domain_slot.split("-")
slot = self.slot_mapping.get(slot, slot)
value = belief_gen[slot_idx][:-1] # remove <EOS>
value = self.tokenizer.decode(value)
if value != "none":
if slot in self.state["belief_state"][domain]["book"].keys():
if self.state["belief_state"][domain]["book"][slot] == "":
action = ["Inform", domain.capitalize(), REF_USR_DA[domain].get(slot, slot), value]
user_action.append(action)
self.state["belief_state"][domain]["book"][slot] = value
elif slot in self.state["belief_state"][domain]["semi"].keys():
if self.state["belief_state"][domain]["semi"][slot] == "":
action = ["Inform", domain.capitalize(), REF_USR_DA[domain].get(slot, slot), value]
user_action.append(action)
self.state["belief_state"][domain]["semi"][slot] = value
self.state["user_action"] = user_action
self.output_action = deepcopy(self.policy.predict(self.state))
model_response = self.nlg.generate(self.output_action)
self.history.append(["sys", model_response])
return model_response
"police":[],
"hospital":["postcode"]
}
with torch.no_grad():
for batch_idx, batch in t:
inputs, contexts, context_lengths, dial_ids = reader.make_input(batch)
batch_size = len(contexts[0])
turns = len(inputs)
nlu.init_session()
dst.init_session()
for turn_idx in range(turns):
context_len = contexts[turn_idx].size(1)
input_action = nlu.predict(inputs[turn_idx]["usr"][0], inputs[turn_idx]["context"][0])
dst.state['user_action'] = input_action
state = dst.update(input_action)
belief = state["belief_state"]
belief_label = inputs[turn_idx]["belief"][0]
# joint_acc_ = 1
# for slot_idx, value in enumerate(belief_label):
# slot = ontology.all_info_slots[slot_idx]
# domain, slot = slot.split("-")
# slot = mapping.get(slot, slot)
# value = value[value != 0].tolist()
# value = vocab.decode(value[:-1])
# if value == "none":
# value = ""
# elif value == "don't care":