def valid_on_epoch(self, epoch, filename, memory):
self.logger.info("Epoch {:02} {} begins validing ...................".format(epoch, self.tag))
self.model.eval()
start_time = time.time()
score = Fscore(self.tag)
datas = SLUDataset.read_file(filename, memory)
for pair in datas:
cnet = pair['cnet']
class_string = pair['label']
gold_classes = SLUDataset.class_info(class_string)
pred_classes = decode_slu(self.model, cnet, memory, self.cuda)
score.update_tp_fp_fn(pred_classes, gold_classes)
fscore = score.output_fscore(self.logger, epoch)
elapsed_time = time.time() - start_time
self.logger.info("Epoch {:02} {} ends validing elapsed_time: {:6.0f}s".format(
epoch, self.tag, elapsed_time)
)
self.logger.info('*****************************************************')
return fscore
def error(opt):
opt.experiment = os.path.join(root_dir, opt.experiment)
opt.load_chkpt = os.path.join(opt.experiment, opt.save_model)
opt.test_file = os.path.join(opt.data_root, opt.test_file)
opt.save_file = os.path.join(opt.experiment, 'error.json')
# Model loading
model = make_model(opt)
chkpt = torch.load(opt.load_chkpt,
map_location=lambda storage, log: storage)
model.load_state_dict(chkpt)
if opt.deviceid >= 0:
model = model.cuda()
print(model)
# ====== *********************** ================
model.eval()
# ===============================================
# decode
print('Decoding ...')
if opt.task == 'act':
datas = ActDataset.read_file(opt.test_file, opt.memory)
elif opt.task == 'slot':
datas = SlotDataset.read_file(opt.test_file, opt.memory)
elif opt.task == 'value':
datas = ValueDataset.read_file(opt.test_file, opt.memory)
elif opt.task == 'slu':
datas = SLUDataset.read_file(opt.test_file, opt.memory)
dic = {'pairs': []}
for pair in datas:
cnet = pair['cnet']
class_string = pair['label']
if opt.task == 'act':
gold_classes = ActDataset.class_info(class_string)
pred_classes = decode_act(model, cnet, opt.memory, opt.cuda)
elif opt.task == 'slot':
gold_classes = SlotDataset.class_info(class_string)
pred_classes = decode_slot(model, cnet, class_string, opt.memory,
opt.cuda)
elif opt.task == 'value':
gold_classes = ValueDataset.class_info(class_string)
pred_classes = decode_value(model, cnet, class_string, opt.memory,
opt.cuda)
elif opt.task == 'slu':
gold_classes = SLUDataset.class_info(class_string)
pred_classes = decode_slu(model, cnet, opt.memory, opt.cuda)
gold_class = ';'.join(sorted(gold_classes))
pred_class = ';'.join(sorted(pred_classes))
if gold_class != pred_class:
pr = {'cnet': cnet, 'label': gold_class, 'pred': pred_class}
dic['pairs'].append(pr)
string = json.dumps(dic, sort_keys=True, indent=4, separators=(',', ';'))
with open(opt.save_file, 'w') as f:
f.write(string)
print('Decode results saved in {}'.format(opt.save_file))
def error(opt):
opt.experiment = os.path.join(root_dir, opt.experiment)
opt.load_chkpt = os.path.join(opt.experiment, opt.save_model)
opt.test_file = os.path.join(opt.data_root, opt.test_file)
opt.save_file = os.path.join(opt.experiment, 'error.info')
# Model loading
model = make_model(opt)
chkpt = torch.load(opt.load_chkpt,
map_location=lambda storage, log: storage)
model.load_state_dict(chkpt)
if opt.deviceid >= 0:
model = model.cuda()
print(model)
# ====== *********************** ================
model.eval()
# ===============================================
# decode
print('Decoding ...')
g = open(opt.save_file, 'w')
if opt.task == 'act':
lines = ActDataset.read_file(opt.test_file)
elif opt.task == 'slot':
lines = SlotDataset.read_file(opt.test_file)
elif opt.task == 'value':
lines = ValueDataset.read_file(opt.test_file)
elif opt.task == 'slu':
lines = SLUDataset.read_file(opt.test_file)
for (utterance, class_string) in lines:
if opt.task == 'act':
gold_classes = ActDataset.class_info(class_string)
pred_classes = decode_act(model, utterance, opt.memory, opt.cuda)
elif opt.task == 'slot':
gold_classes = SlotDataset.class_info(class_string)
pred_classes = decode_slot(model, utterance, class_string,
opt.memory, opt.cuda)
elif opt.task == 'value':
gold_classes = ValueDataset.class_info(class_string)
pred_classes = decode_value(model, utterance, class_string,
opt.memory, opt.cuda)
elif opt.task == 'slu':
gold_classes = SLUDataset.class_info(class_string)
pred_classes = decode_slu(model, utterance, opt.memory, opt.cuda)
gold_class = ';'.join(sorted(gold_classes))
pred_class = ';'.join(sorted(pred_classes))
if gold_class != pred_class:
g.write('{}\t<=>\t{}\t<=>\t{}\n'.format(utterance, gold_class,
pred_class))
g.close()
print('Decode results saved in {}'.format(opt.save_file))
def decode_slu(model, cnet, memory, cuda):
result = process_cn_example(cnet, memory['enc2idx'])
if result is None:
return []
data, lengths, extra_zeros, enc_batch_extend_vocab_idx, oov_list = \
SLUDataset.data_info(cnet, memory, cuda)
act_slot_values = []
# Model processing
## encoder
outputs, hiddens = model.encoder(data, lengths)
h_T = hiddens[0]
## act prediction
act_scores = model.act_stc(h_T)
act_scores = act_scores.data.cpu().view(-1, ).numpy()
pred_acts = [i for i, p in enumerate(act_scores) if p > 0.5]
act_pairs = [(i, memory['idx2act'][i]) for i in pred_acts]
remain_acts = []
for act in act_pairs:
if act[1] == 'pad':
continue
elif act[1] in memory['single_acts']:
act_slot_values.append(act[1])
else:
remain_acts.append(act)
if len(remain_acts) == 0:
return act_slot_values
## slot prediction
remain_act_slots = []
for act in remain_acts:
act_input = torch.tensor([act[0]]).view(1, 1)
if cuda:
act_input = act_input.cuda()
slot_scores = model.slot_predict(h_T, act_input)
slot_scores = slot_scores.data.cpu().view(-1, ).numpy()
pred_slots = [i for i, p in enumerate(slot_scores) if p > 0.5]
slot_pairs = [(i, memory['idx2slot'][i]) for i in pred_slots]
if act[1] in memory['double_acts']:
for slot in slot_pairs:
act_slot_values.append('-'.join([act[1], slot[1]]))
else:
for slot in slot_pairs:
if slot[1] != 'pad':
remain_act_slots.append(list(zip(act, slot)))
if len(remain_act_slots) == 0:
return act_slot_values
## value decoder
s_decoder = model.enc_to_dec(hiddens)
for act_slot in remain_act_slots:
s_t_1 = s_decoder
act_slot_ids = torch.tensor(act_slot[0]).view(1, 2)
y_t = torch.tensor([Constants.BOS]).view(1, 1)
if cuda:
y_t = y_t.cuda()
act_slot_ids = act_slot_ids.cuda()
value_ids = beam_search(model.value_decoder, act_slot_ids, extra_zeros,
enc_batch_extend_vocab_idx, s_decoder, outputs,
lengths, len(memory['dec2idx']), cuda)[1:-1]
value_lis = []
for vid in value_ids:
if vid < len(memory['idx2dec']):
value_lis.append(memory['idx2dec'][vid])
else:
value_lis.append(oov_list[vid - len(memory['idx2dec'])])
values = ' '.join(value_lis)
act_slot_values.append('-'.join(list(act_slot[1]) + [values]))
return act_slot_values
def train(opt):
opt.experiment = os.path.join(root_dir, opt.experiment)
if not os.path.exists(opt.experiment):
os.makedirs(opt.experiment)
opt.save_model = os.path.join(opt.experiment, opt.save_model)
opt.log_path = os.path.join(opt.experiment, 'log.train')
opt.logger = make_logger(opt.log_path)
# memory info
print("encoder word2idx number: {}".format(opt.enc_word_vocab_size))
print("decoder word2idx number: {}".format(opt.dec_word_vocab_size))
print("act2idx number: {}".format(opt.act_vocab_size))
print("slot2idx number: {}".format(opt.slot_vocab_size))
# Model definition
model = make_model(opt)
if opt.load_word_emb:
emb = read_emb(opt.memory['enc2idx'])
model.enc_word_emb.init_weight_from_pre_emb(emb, opt.fix_word_emb)
if opt.load_class_emb:
emb = opt.memory['act_emb']
model.act_emb.init_weight_from_pre_emb(emb, opt.fix_class_emb)
emb = opt.memory['slot_emb']
model.slot_emb.init_weight_from_pre_emb(emb, opt.fix_class_emb)
if opt.share_param:
#model.value_decoder.outlin.weight.data = model.word_emb.embedding.weight.data
#model.value_decoder.outlin.weight.requires_grad = model.word_emb.embedding.weight.requires_grad
model.act_stc.lin.weight.data = model.act_emb.embedding.weight.data
model.act_stc.lin.weight.requires_grad = model.act_emb.embedding.weight.requires_grad
model.slot_stc.lin.weight.data = model.slot_emb.embedding.weight.data
model.slot_stc.lin.weight.requires_grad = model.slot_emb.embedding.weight.requires_grad
if opt.cuda:
model = model.cuda()
print(model)
# optimizer details
optimizer = Optim(opt.optim, opt.lr, max_grad_norm=opt.max_norm)
optimizer.set_parameters(model.named_parameters())
print("training parameters number: {}".format(len(optimizer.params)))
# loss definition
#stc_criterion = MaskedBCELoss(opt.cuda)
stc_criterion = nn.BCELoss(reduction='sum')
nll_criterion = nn.NLLLoss(reduction='sum')
if opt.cuda:
stc_criterion = stc_criterion.cuda()
nll_criterion = nll_criterion.cuda()
# training procedure
if opt.task == 'slu':
data_iter = SLUDataset(opt.data_root + 'train', opt.memory, opt.cuda,
True)
trainer = SLUTrainer(model, (stc_criterion, nll_criterion),
optimizer,
opt.logger,
cuda=opt.cuda)
elif opt.task == 'act':
data_iter = ActDataset(opt.data_root + 'train', opt.memory, opt.cuda,
True)
trainer = ActTrainer(model,
stc_criterion,
optimizer,
opt.logger,
cuda=opt.cuda)
elif opt.task == 'slot':
data_iter = SlotDataset(opt.data_root + 'train', opt.memory, opt.cuda,
True)
trainer = SlotTrainer(model,
stc_criterion,
optimizer,
opt.logger,
cuda=opt.cuda)
elif opt.task == 'value':
data_iter = ValueDataset(opt.data_root + 'train', opt.memory, opt.cuda,
True)
trainer = ValueTrainer(model,
nll_criterion,
optimizer,
opt.logger,
cuda=opt.cuda)
trainer.train(opt.epochs, opt.batch_size, opt.memory, data_iter,
opt.data_root + 'valid', opt.save_model)