def stat(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.log_path = os.path.join(opt.experiment, 'stat.info')
opt.logger = make_logger(opt.log_path)
# 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.cuda:
model = model.cuda()
# ====== *********************** ================
model.eval()
# ===============================================
# training procedure
trainers = []
trainers.append(ActTrainer(model, None, None, opt.logger, cuda=opt.cuda))
trainers.append(SlotTrainer(model, None, None, opt.logger, cuda=opt.cuda))
trainers.append(ValueTrainer(model, None, None, opt.logger, cuda=opt.cuda))
trainers.append(
SLUTrainer(model, (None, None), None, opt.logger, cuda=opt.cuda))
for trainer in trainers:
metric = trainer.valid_on_epoch(0, opt.test_file, opt.memory)
print('Stat results saved in {}'.format(opt.log_path))
def train(opt):
# basics definition
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)
bert_tokenizer, bert_model = make_bert()
# dataIter definition
class2idx = build_class_vocab(opt.data_root + 'class.all')
opt.class_size = len(class2idx)
train_iter = BertIter4STC(opt.data_root + 'train', bert_tokenizer,
class2idx, opt.batch_size, opt.cuda, True)
valid_iter = BertIter4STC(opt.data_root + 'valid', bert_tokenizer,
class2idx, opt.batch_size, opt.cuda, False)
# model definition
model = make_model(opt, bert_model)
# criterion definition
criterion = nn.BCELoss(reduction='sum')
if opt.cuda:
criterion = criterion.cuda()
# optimizer definition
if opt.fix_bert:
for (name, parameter) in model.bert.named_parameters():
parameter.requires_grad = False
if opt.optim == 'bert':
params = list(
filter(lambda x: x[1].requires_grad == True,
model.named_parameters()))
print('Trainable parameter number: {}'.format(len(params)))
print('Trainer: bert')
no_decay = ['bias', 'gamma', 'beta']
grouped_params = [{
'params': [p for n, p in params if n not in no_decay],
'weight_decay_rate': 0.01
}, {
'params': [p for n, p in params if n in no_decay],
'weight_decay_rate': 0.0
}]
optimizer = BertAdam(grouped_params,
opt.lr,
warmup=0.1,
t_total=len(train_iter) * opt.epochs)
else:
optimizer = Optim(opt.optim, opt.lr, max_grad_norm=opt.max_norm)
optimizer.set_parameters(model.named_parameters())
print('Trainable parameter number: {}'.format(len(optimizer.params)))
# training procedure
trainer = BertTrainer4STC(model, criterion, optimizer, opt.logger)
trainer.train(opt.epochs, train_iter, valid_iter, opt.save_model)
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)
if opt.load_model is not None:
opt.load_model = os.path.join(opt.experiment, opt.load_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))
# Model definition
model = make_model(opt)
if opt.load_word_emb:
if opt.enlarge_word_vocab:
enc_emb = opt.memory['word2idx_w_glove_emb']
else:
enc_emb = opt.memory['word2idx_emb']
dec_emb = opt.memory['word2idx_emb']
model.enc_word_emb.init_weight_from_pre_emb(enc_emb, opt.fix_word_emb)
model.dec_word_emb.init_weight_from_pre_emb(dec_emb, opt.fix_word_emb)
if opt.enc_word_vocab_size == opt.dec_word_vocab_size:
model.dec_word_emb.embedding.weight.data = model.enc_word_emb.embedding.weight.data
model.dec_word_emb.embedding.weight.requires_grad = model.enc_word_emb.embedding.weight.requires_grad
if opt.load_model is not None:
chkpt = torch.load(opt.load_model, map_location = lambda storage, log: storage)
model.load_state_dict(chkpt)
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)))
nll_criterion = nn.NLLLoss(reduction='sum')
if opt.cuda:
nll_criterion = nll_criterion.cuda()
# training procedure
train_iter = DADataset(opt.data_root + opt.train_file, opt.memory, opt.cuda, True)
valid_iter = DADataset(opt.data_root + opt.valid_file, opt.memory, opt.cuda, False)
trainer = DATrainer(model, nll_criterion, optimizer, opt.logger, cuda=opt.cuda)
trainer.train(opt.epochs, opt.batch_size, train_iter, valid_iter, opt.save_model)
def train():
# Initialization
opt = parse_args()
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)
# Data iterator definition
train_iter = DataLoader(opt.train_file,
opt.word2idx,
opt.batch_size,
opt.cuda,
epoch_shuffle=True)
valid_iter = DataLoader(opt.valid_file,
opt.word2idx,
opt.batch_size,
opt.cuda,
epoch_shuffle=False)
# Model definition
model = make_model(len(opt.word2idx))
if opt.gpuid >= 0:
model = model.cuda()
#print(model)
tally_parameters(model)
optimizer = Optim('Adam', 0.001, max_grad_norm=5)
optimizer.set_parameters(model.named_parameters())
criterion = nn.NLLLoss(reduction='sum')
if opt.gpuid >= 0:
criterion = criterion.cuda()
trainer = Trainer(model, criterion, optimizer, opt.logger, opt.cuda)
trainer.train(opt.epochs, train_iter, valid_iter, opt.save_model)
def train(opt):
# basics definition
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)
# dataIter definition
train_iter = OneBestIter4STC(opt.data_root+'train', opt.word2idx, opt.class2idx,
opt.batch_size, opt.cuda, True)
valid_iter = OneBestIter4STC(opt.data_root+'valid', opt.word2idx, opt.class2idx,
opt.batch_size, opt.cuda, False)
# model definition
model = make_model(opt)
if opt.load_emb:
emb = read_emb(opt.word2idx)
model.emb.init_weight_from_pre_emb(emb, opt.fix_emb)
print(model)
# criterion definition
criterion = nn.BCELoss(reduction='sum')
if opt.cuda:
criterion = criterion.cuda()
# optimizer definition
optimizer = Optim(opt.optim, opt.lr, max_grad_norm=opt.max_norm)
optimizer.set_parameters(model.named_parameters())
print('Trainable parameter number: {}'.format(len(optimizer.params)))
# training procedure
trainer = OneBestTrainer4STC(model, criterion, optimizer, opt.logger)
trainer.train(opt.epochs, train_iter, valid_iter, opt.save_model)
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)