def main(args):
def worker_init_fn(worker_id):
np.random.seed(args.random_seed + worker_id)
n_gpu = 0
if torch.cuda.is_available():
n_gpu = torch.cuda.device_count()
np.random.seed(args.random_seed)
random.seed(args.random_seed)
rng = random.Random(args.random_seed)
torch.manual_seed(args.random_seed)
if n_gpu > 0:
torch.cuda.manual_seed(args.random_seed)
torch.cuda.manual_seed_all(args.random_seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
if not os.path.exists(args.save_dir):
os.mkdir(args.save_dir)
ontology = json.load(open(args.ontology_data))
slot_meta, ontology = make_slot_meta(ontology)
op2id = OP_SET[args.op_code]
print(op2id)
tokenizer = BertTokenizer(args.vocab_path, do_lower_case=True)
train_data_raw = prepare_dataset(data_path=args.train_data_path,
tokenizer=tokenizer,
slot_meta=slot_meta,
n_history=args.n_history,
max_seq_length=args.max_seq_length,
op_code=args.op_code)
train_data = MultiWozDataset(train_data_raw,
tokenizer,
slot_meta,
args.max_seq_length,
rng,
ontology,
args.word_dropout,
args.shuffle_state,
args.shuffle_p)
print("# train examples %d" % len(train_data_raw))
dev_data_raw = prepare_dataset(data_path=args.dev_data_path,
tokenizer=tokenizer,
slot_meta=slot_meta,
n_history=args.n_history,
max_seq_length=args.max_seq_length,
op_code=args.op_code)
print("# dev examples %d" % len(dev_data_raw))
test_data_raw = prepare_dataset(data_path=args.test_data_path,
tokenizer=tokenizer,
slot_meta=slot_meta,
n_history=args.n_history,
max_seq_length=args.max_seq_length,
op_code=args.op_code)
print("# test examples %d" % len(test_data_raw))
model_config = BertConfig.from_json_file(args.bert_config_path)
model_config.dropout = args.dropout
model_config.attention_probs_dropout_prob = args.attention_probs_dropout_prob
model_config.hidden_dropout_prob = args.hidden_dropout_prob
model = SomDST(model_config, len(op2id), len(domain2id), op2id['update'], args.exclude_domain)
if not os.path.exists(args.bert_ckpt_path):
args.bert_ckpt_path = download_ckpt(args.bert_ckpt_path, args.bert_config_path, 'assets')
ckpt = torch.load(args.bert_ckpt_path, map_location='cpu')
model.encoder.bert.load_state_dict(ckpt)
# re-initialize added special tokens ([SLOT], [NULL], [EOS])
model.encoder.bert.embeddings.word_embeddings.weight.data[1].normal_(mean=0.0, std=0.02)
model.encoder.bert.embeddings.word_embeddings.weight.data[2].normal_(mean=0.0, std=0.02)
model.encoder.bert.embeddings.word_embeddings.weight.data[3].normal_(mean=0.0, std=0.02)
model.to(device)
num_train_steps = int(len(train_data_raw) / args.batch_size * args.n_epochs)
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
enc_param_optimizer = list(model.encoder.named_parameters())
enc_optimizer_grouped_parameters = [
{'params': [p for n, p in enc_param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in enc_param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
enc_optimizer = AdamW(enc_optimizer_grouped_parameters, lr=args.enc_lr)
enc_scheduler = WarmupLinearSchedule(enc_optimizer, int(num_train_steps * args.enc_warmup),
t_total=num_train_steps)
dec_param_optimizer = list(model.decoder.parameters())
dec_optimizer = AdamW(dec_param_optimizer, lr=args.dec_lr)
dec_scheduler = WarmupLinearSchedule(dec_optimizer, int(num_train_steps * args.dec_warmup),
t_total=num_train_steps)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.batch_size,
collate_fn=train_data.collate_fn,
num_workers=args.num_workers,
worker_init_fn=worker_init_fn)
loss_fnc = nn.CrossEntropyLoss()
best_score = {'epoch': 0, 'joint_acc': 0, 'op_acc': 0, 'final_slot_f1': 0}
for epoch in range(args.n_epochs):
batch_loss = []
model.train()
for step, batch in enumerate(train_dataloader):
batch = [b.to(device) if not isinstance(b, int) else b for b in batch]
input_ids, input_mask, segment_ids, state_position_ids, op_ids,\
domain_ids, gen_ids, max_value, max_update = batch
if rng.random() < args.decoder_teacher_forcing: # teacher forcing
teacher = gen_ids
else:
teacher = None
domain_scores, state_scores, gen_scores = model(input_ids=input_ids,
token_type_ids=segment_ids,
state_positions=state_position_ids,
attention_mask=input_mask,
max_value=max_value,
op_ids=op_ids,
max_update=max_update,
teacher=teacher)
loss_s = loss_fnc(state_scores.view(-1, len(op2id)), op_ids.view(-1))
loss_g = masked_cross_entropy_for_value(gen_scores.contiguous(),
gen_ids.contiguous(),
tokenizer.vocab['[PAD]'])
loss = loss_s + loss_g
if args.exclude_domain is not True:
loss_d = loss_fnc(domain_scores.view(-1, len(domain2id)), domain_ids.view(-1))
loss = loss + loss_d
batch_loss.append(loss.item())
loss.backward()
enc_optimizer.step()
enc_scheduler.step()
dec_optimizer.step()
dec_scheduler.step()
model.zero_grad()
if step % 100 == 0:
if args.exclude_domain is not True:
print("[%d/%d] [%d/%d] mean_loss : %.3f, state_loss : %.3f, gen_loss : %.3f, dom_loss : %.3f" \
% (epoch+1, args.n_epochs, step,
len(train_dataloader), np.mean(batch_loss),
loss_s.item(), loss_g.item(), loss_d.item()))
else:
print("[%d/%d] [%d/%d] mean_loss : %.3f, state_loss : %.3f, gen_loss : %.3f" \
% (epoch+1, args.n_epochs, step,
len(train_dataloader), np.mean(batch_loss),
loss_s.item(), loss_g.item()))
batch_loss = []
if (epoch+1) % args.eval_epoch == 0:
eval_res = model_evaluation(model, dev_data_raw, tokenizer, slot_meta, epoch+1, args.op_code)
if eval_res['joint_acc'] > best_score['joint_acc']:
best_score = eval_res
model_to_save = model.module if hasattr(model, 'module') else model
save_path = os.path.join(args.save_dir, 'model_best.bin')
torch.save(model_to_save.state_dict(), save_path)
print("Best Score : ", best_score)
print("\n")
print("Test using best model...")
best_epoch = best_score['epoch']
ckpt_path = os.path.join(args.save_dir, 'model_best.bin')
model = SomDST(model_config, len(op2id), len(domain2id), op2id['update'], args.exclude_domain)
ckpt = torch.load(ckpt_path, map_location='cpu')
model.load_state_dict(ckpt)
model.to(device)
model_evaluation(model, test_data_raw, tokenizer, slot_meta, best_epoch, args.op_code,
is_gt_op=False, is_gt_p_state=False, is_gt_gen=False)
model_evaluation(model, test_data_raw, tokenizer, slot_meta, best_epoch, args.op_code,
is_gt_op=False, is_gt_p_state=False, is_gt_gen=True)
model_evaluation(model, test_data_raw, tokenizer, slot_meta, best_epoch, args.op_code,
is_gt_op=False, is_gt_p_state=True, is_gt_gen=False)
model_evaluation(model, test_data_raw, tokenizer, slot_meta, best_epoch, args.op_code,
is_gt_op=False, is_gt_p_state=True, is_gt_gen=True)
model_evaluation(model, test_data_raw, tokenizer, slot_meta, best_epoch, args.op_code,
is_gt_op=True, is_gt_p_state=False, is_gt_gen=False)
model_evaluation(model, test_data_raw, tokenizer, slot_meta, best_epoch, args.op_code,
is_gt_op=True, is_gt_p_state=True, is_gt_gen=False)
model_evaluation(model, test_data_raw, tokenizer, slot_meta, best_epoch, args.op_code,
is_gt_op=True, is_gt_p_state=False, is_gt_gen=True)
model_evaluation(model, test_data_raw, tokenizer, slot_meta, best_epoch, args.op_code,
is_gt_op=True, is_gt_p_state=True, is_gt_gen=True)
def main(args):
assert args.use_one_optim is True
if args.use_cls_only:
args.no_dial = True
print("### use_cls_only: {:}".format(args.use_cls_only))
print("### no_dial: {:}".format(args.no_dial))
if args.recover_e > 0:
raise NotImplementedError("This option is from my oldest code version. "
"I have not checked it for this code version.")
if not os.path.exists(args.save_dir):
os.mkdir(args.save_dir)
print("### mkdir {:}".format(args.save_dir))
def worker_init_fn(worker_id):
np.random.seed(args.random_seed + worker_id)
n_gpu = 0
if torch.cuda.is_available() and (not args.use_cpu):
n_gpu = torch.cuda.device_count()
device = torch.device('cuda')
print("### Device: {:}".format(device))
else:
print("### Use CPU (Debugging)")
device = torch.device("cpu")
if args.random_seed < 0:
print("### Pick a random seed")
args.random_seed = random.sample(list(range(0, 100000)), 1)[0]
print("### Random Seed: {:}".format(args.random_seed))
np.random.seed(args.random_seed)
random.seed(args.random_seed)
rng = random.Random(args.random_seed)
torch.manual_seed(args.random_seed)
if n_gpu > 0:
if args.random_seed >= 0:
torch.cuda.manual_seed(args.random_seed)
torch.cuda.manual_seed_all(args.random_seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
if not os.path.exists(args.save_dir):
os.mkdir(args.save_dir)
ontology = json.load(open(args.ontology_data))
slot_meta, ontology = make_slot_meta(ontology)
op2id = OP_SET[args.op_code]
print(op2id)
tokenizer = BertTokenizer(args.vocab_path, do_lower_case=True)
train_path = os.path.join(args.data_root, "train.pt")
dev_path = os.path.join(args.data_root, "dev.pt")
test_path = os.path.join(args.data_root, "test.pt")
if not os.path.exists(test_path):
test_data_raw = prepare_dataset(data_path=args.test_data_path,
tokenizer=tokenizer,
slot_meta=slot_meta,
n_history=args.n_history,
max_seq_length=args.max_seq_length,
op_code=args.op_code)
torch.save(test_data_raw, test_path)
else:
test_data_raw = torch.load(test_path)
print("# test examples %d" % len(test_data_raw))
if not os.path.exists(train_path):
train_data_raw = prepare_dataset(data_path=args.train_data_path,
tokenizer=tokenizer,
slot_meta=slot_meta,
n_history=args.n_history,
max_seq_length=args.max_seq_length,
op_code=args.op_code)
torch.save(train_data_raw, train_path)
else:
train_data_raw = torch.load(train_path)
train_data = MultiWozDataset(train_data_raw,
tokenizer,
slot_meta,
args.max_seq_length,
rng,
ontology,
args.word_dropout,
args.shuffle_state,
args.shuffle_p, pad_id=tokenizer.convert_tokens_to_ids(['[PAD]'])[0],
slot_id=tokenizer.convert_tokens_to_ids(['[SLOT]'])[0],
decoder_teacher_forcing=args.decoder_teacher_forcing,
use_full_slot=args.use_full_slot,
use_dt_only=args.use_dt_only, no_dial=args.no_dial,
use_cls_only=args.use_cls_only)
print("# train examples %d" % len(train_data_raw))
if not os.path.exists(dev_path):
dev_data_raw = prepare_dataset(data_path=args.dev_data_path,
tokenizer=tokenizer,
slot_meta=slot_meta,
n_history=args.n_history,
max_seq_length=args.max_seq_length,
op_code=args.op_code)
torch.save(dev_data_raw, dev_path)
else:
dev_data_raw = torch.load(dev_path)
print("# dev examples %d" % len(dev_data_raw))
model_config = BertConfig.from_json_file(args.bert_config_path)
model_config.dropout = args.dropout
model_config.attention_probs_dropout_prob = args.attention_probs_dropout_prob
model_config.hidden_dropout_prob = args.hidden_dropout_prob
type_vocab_size = 4
dec_config = args
model = TransformerDST(model_config, dec_config, len(op2id), len(domain2id),
op2id['update'],
tokenizer.convert_tokens_to_ids(['[MASK]'])[0],
tokenizer.convert_tokens_to_ids(['[SEP]'])[0],
tokenizer.convert_tokens_to_ids(['[PAD]'])[0],
tokenizer.convert_tokens_to_ids(['-'])[0],
type_vocab_size, args.exclude_domain)
if not os.path.exists(args.bert_ckpt_path):
args.bert_ckpt_path = download_ckpt(args.bert_ckpt_path, args.bert_config_path, 'assets')
state_dict = torch.load(args.bert_ckpt_path, map_location='cpu')
_k = 'embeddings.token_type_embeddings.weight'
print("config.type_vocab_size != state_dict[bert.embeddings.token_type_embeddings.weight] ({0} != {1})".format(
type_vocab_size, state_dict[_k].shape[0]))
state_dict[_k].resize_(
type_vocab_size, state_dict[_k].shape[1])
state_dict[_k].data[2, :].copy_(state_dict[_k].data[0, :])
state_dict[_k].data[3, :].copy_(state_dict[_k].data[0, :])
model.bert.load_state_dict(state_dict)
print("\n### Done Load BERT")
sys.stdout.flush()
# re-initialize added special tokens ([SLOT], [NULL], [EOS])
model.bert.embeddings.word_embeddings.weight.data[1].normal_(mean=0.0, std=0.02)
model.bert.embeddings.word_embeddings.weight.data[2].normal_(mean=0.0, std=0.02)
model.bert.embeddings.word_embeddings.weight.data[3].normal_(mean=0.0, std=0.02)
# re-initialize seg-2, seg-3
model.bert.embeddings.token_type_embeddings.weight.data[2].normal_(mean=0.0, std=0.02)
model.bert.embeddings.token_type_embeddings.weight.data[3].normal_(mean=0.0, std=0.02)
model.to(device)
num_train_steps = int(len(train_data_raw) / args.batch_size * args.n_epochs)
if args.use_one_optim:
print("### Use One Optim")
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(
nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(
nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.enc_lr)
scheduler = WarmupLinearSchedule(optimizer, int(num_train_steps * args.enc_warmup),
t_total=num_train_steps)
else:
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
enc_param_optimizer = list(model.bert.named_parameters()) # TODO: For BERT only
print('### Optim BERT: {:}'.format(len(enc_param_optimizer)))
enc_optimizer_grouped_parameters = [
{'params': [p for n, p in enc_param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in enc_param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
enc_optimizer = AdamW(enc_optimizer_grouped_parameters, lr=args.enc_lr)
enc_scheduler = WarmupLinearSchedule(enc_optimizer, int(num_train_steps * args.enc_warmup),
t_total=num_train_steps)
dec_param_optimizer = list(model.named_parameters()) # TODO: For other parameters
print('### Optim All: {:}'.format(len(dec_param_optimizer)))
dec_param_optimizer = [p for (n, p) in dec_param_optimizer if 'bert' not in n]
print('### Optim OTH: {:}'.format(len(dec_param_optimizer)))
dec_optimizer = AdamW(dec_param_optimizer, lr=args.dec_lr)
dec_scheduler = WarmupLinearSchedule(dec_optimizer, int(num_train_steps * args.dec_warmup),
t_total=num_train_steps)
if args.recover_e > 0:
model_recover, enc_recover, dec_recover = load(args, str(args.recover_e))
print("### Recover Model E{:}".format(args.recover_e))
sys.stdout.flush()
model.load_state_dict(model_recover)
print("### Recover Optim E{:}".format(args.recover_e))
sys.stdout.flush()
enc_optimizer.load_state_dict(enc_recover)
dec_optimizer.load_state_dict(dec_optimizer)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.batch_size,
collate_fn=train_data.collate_fn,
num_workers=args.num_workers,
worker_init_fn=worker_init_fn)
loss_fnc = nn.CrossEntropyLoss()
best_score = {'epoch': 0, 'joint_acc': 0, 'op_acc': 0, 'final_slot_f1': 0}
start_time = time.time()
for epoch in range(args.n_epochs):
batch_loss = []
model.train()
for step, batch in enumerate(train_dataloader):
batch = [b.to(device) if (not isinstance(b, int)) and (not isinstance(b, dict) and (not isinstance(b, list)) and (not isinstance(b, np.ndarray))) else b for b in batch]
input_ids_p, segment_ids_p, input_mask_p, \
state_position_ids, op_ids, domain_ids, input_ids_g, segment_ids_g, position_ids_g, input_mask_g, \
masked_pos, masked_weights, lm_label_ids, id_n_map, gen_max_len, n_total_pred = batch
domain_scores, state_scores, loss_g = model(input_ids_p, segment_ids_p, input_mask_p, state_position_ids,
input_ids_g, segment_ids_g, position_ids_g, input_mask_g,
masked_pos, masked_weights, lm_label_ids, id_n_map, gen_max_len, only_pred_op=args.only_pred_op, n_gpu=n_gpu)
if n_total_pred > 0:
loss_g = loss_g.sum() / n_total_pred
else:
loss_g = 0
loss_s = loss_fnc(state_scores.view(-1, len(op2id)), op_ids.view(-1))
if args.only_pred_op:
loss = loss_s
else:
loss = loss_s + loss_g
if args.exclude_domain is not True:
loss_d = loss_fnc(domain_scores.view(-1, len(domain2id)), domain_ids.view(-1))
loss = loss + loss_d
batch_loss.append(loss.item())
loss.backward()
if args.use_one_optim:
optimizer.step()
scheduler.step()
else:
enc_optimizer.step()
enc_scheduler.step()
dec_optimizer.step()
dec_scheduler.step()
model.zero_grad()
if step % 100 == 0:
try:
loss_g = loss_g.item()
except AttributeError:
loss_g = loss_g
if args.exclude_domain is not True:
print("time %.1f min, [%d/%d] [%d/%d] mean_loss : %.3f, state_loss : %.3f, gen_loss : %.3f, dom_loss : %.3f" \
% ((time.time()-start_time)/60, epoch+1, args.n_epochs, step,
len(train_dataloader), np.mean(batch_loss),
loss_s.item(), loss_g, loss_d.item()))
else:
print("time %.1f min, [%d/%d] [%d/%d] mean_loss : %.3f, state_loss : %.3f, gen_loss : %.3f" \
% ((time.time()-start_time)/60, epoch+1, args.n_epochs, step,
len(train_dataloader), np.mean(batch_loss),
loss_s.item(), loss_g))
sys.stdout.flush()
batch_loss = []
if args.use_one_optim:
save(args, epoch + 1, model, optimizer)
else:
save(args, epoch + 1, model, enc_optimizer, dec_optimizer)
if ((epoch+1) % args.eval_epoch == 0) and (epoch+1 >= 8):
eval_res = model_evaluation(model, dev_data_raw, tokenizer, slot_meta, epoch+1, args.op_code,
use_full_slot=args.use_full_slot, use_dt_only=args.use_dt_only, no_dial=args.no_dial, use_cls_only=args.use_cls_only, n_gpu=n_gpu)
print("### Epoch {:} Score : ".format(epoch+1), eval_res)
if eval_res['joint_acc'] > best_score['joint_acc']:
best_score = eval_res
print("### Best Joint Acc: {:} ###".format(best_score['joint_acc']))
print('\n')
if epoch+1 >= 8: # To speed up
eval_res_test = model_evaluation(model, test_data_raw, tokenizer, slot_meta, epoch + 1, args.op_code,
use_full_slot=args.use_full_slot, use_dt_only=args.use_dt_only, no_dial=args.no_dial, use_cls_only=args.use_cls_only, n_gpu=n_gpu)
print("### Epoch {:} Test Score : ".format(epoch + 1), eval_res_test)
def main(args):
def worker_init_fn(worker_id):
np.random.seed(args.random_seed + worker_id)
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
# logger
logger_file_name = args.save_dir.split('/')[1]
fileHandler = logging.FileHandler(
os.path.join(args.save_dir, "%s.txt" % (logger_file_name)))
logger.addHandler(fileHandler)
logger.info(args)
# cuda setup
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
logger.info("device: {}".format(device))
# set random seed
np.random.seed(args.random_seed)
random.seed(args.random_seed)
torch.manual_seed(args.random_seed)
if device == "cuda":
torch.cuda.manual_seed(args.random_seed)
torch.cuda.manual_seed_all(args.random_seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
#******************************************************
# load data
#******************************************************
processor = Processor(args)
slot_meta = processor.slot_meta
label_list = processor.label_list
num_labels = [len(labels) for labels in label_list]
logger.info(slot_meta)
tokenizer = BertTokenizer.from_pretrained(args.pretrained_model)
train_data_raw = processor.get_train_instances(args.data_dir, tokenizer)
print("# train examples %d" % len(train_data_raw))
dev_data_raw = processor.get_dev_instances(args.data_dir, tokenizer)
print("# dev examples %d" % len(dev_data_raw))
test_data_raw = processor.get_test_instances(args.data_dir, tokenizer)
print("# test examples %d" % len(test_data_raw))
logger.info("Data loaded!")
train_data = MultiWozDataset(train_data_raw,
tokenizer,
word_dropout=args.word_dropout)
num_train_steps = int(
len(train_data_raw) / args.train_batch_size * args.n_epochs)
logger.info("***** Run training *****")
logger.info(" Num examples = %d", len(train_data_raw))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size,
collate_fn=train_data.collate_fn,
num_workers=args.num_workers,
worker_init_fn=worker_init_fn)
#******************************************************
# build model
#******************************************************
## Initialize slot and value embeddings
sv_encoder = UtteranceEncoding.from_pretrained(args.pretrained_model)
for p in sv_encoder.bert.parameters():
p.requires_grad = False
new_label_list, slot_value_pos = combine_slot_values(
slot_meta, label_list) # without slot head
logger.info(slot_value_pos)
slot_lookup = get_label_lookup_from_first_token(slot_meta, tokenizer,
sv_encoder, device)
value_lookup = get_label_lookup_from_first_token(new_label_list, tokenizer,
sv_encoder, device)
model = BeliefTracker(args, slot_lookup, value_lookup, num_labels,
slot_value_pos, device)
model.to(device)
## prepare optimizer
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
enc_param_optimizer = list(model.encoder.named_parameters())
enc_optimizer_grouped_parameters = [{
'params': [
p for n, p in enc_param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
0.01
}, {
'params':
[p for n, p in enc_param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
enc_optimizer = AdamW(enc_optimizer_grouped_parameters, lr=args.enc_lr)
enc_scheduler = get_linear_schedule_with_warmup(
enc_optimizer, int(num_train_steps * args.enc_warmup), num_train_steps)
dec_param_optimizer = list(model.decoder.parameters())
dec_optimizer = AdamW(dec_param_optimizer, lr=args.dec_lr)
dec_scheduler = get_linear_schedule_with_warmup(
dec_optimizer, int(num_train_steps * args.dec_warmup), num_train_steps)
logger.info(enc_optimizer)
logger.info(dec_optimizer)
#******************************************************
# training
#******************************************************
logger.info("Training...")
best_loss = None
best_acc = None
last_update = None
for epoch in trange(int(args.n_epochs), desc="Epoch"):
batch_loss = []
batch_acc = []
for step, batch in enumerate(tqdm(train_dataloader)):
model.train()
batch = [b.to(device) if b is not None else b for b in batch]
input_ids, segment_ids, input_mask, label_ids = batch
# forward
loss, _, acc, _, _ = model(input_ids=input_ids,
attention_mask=input_mask,
token_type_ids=segment_ids,
labels=label_ids)
loss.backward()
enc_optimizer.step()
enc_scheduler.step()
dec_optimizer.step()
dec_scheduler.step()
model.zero_grad()
batch_loss.append(loss.item())
batch_acc.append(acc)
if step % 300 == 0:
print("[%d/%d] [%d/%d] mean_loss: %.6f, mean_joint_acc: %.6f" % \
(epoch+1, args.n_epochs, step, len(train_dataloader), np.mean(batch_loss), np.mean(batch_acc)))
batch_loss = []
batch_acc = []
if epoch > args.n_epochs / 2 and step > 0 and step % args.eval_step == 0:
eval_res = model_evaluation(model, dev_data_raw, tokenizer,
slot_meta, label_list,
epoch * 10 + step / args.eval_step)
if last_update is None or best_loss > eval_res['loss']:
best_loss = eval_res['loss']
save_path = os.path.join(args.save_dir,
'model_best_loss.bin')
torch.save(model.state_dict(), save_path)
print("Best Loss : ", best_loss)
print("\n")
if last_update is None or best_acc < eval_res['joint_acc']:
best_acc = eval_res['joint_acc']
save_path = os.path.join(args.save_dir,
'model_best_acc.bin')
torch.save(model.state_dict(), save_path)
print("Best Acc : ", best_acc)
print("\n")
logger.info("*** Step=%d, Dev Loss=%.6f, Dev Acc=%.6f, Dev Turn Acc=%.6f, Best Loss=%.6f, Best Acc=%.6f ***" % \
(step, eval_res['loss'], eval_res['joint_acc'], eval_res['joint_turn_acc'], best_loss, best_acc))
if epoch > args.n_epochs / 2 and step > 0 and step % args.eval_step == 0:
eval_res = model_evaluation(model, test_data_raw, tokenizer, slot_meta, \
label_list, epoch*10+step/args.eval_step)
logger.info("*** Step=%d, Tes Loss=%.6f, Tes Acc=%.6f, Tes Turn Acc=%.6f, Best Loss=%.6f, Best Acc=%.6f ***" % \
(step, eval_res['loss'], eval_res['joint_acc'], eval_res['joint_turn_acc'], best_loss, best_acc))
if (epoch + 1) % args.eval_epoch == 0:
eval_res = model_evaluation(model, dev_data_raw, tokenizer,
slot_meta, label_list, epoch + 1)
if last_update is None or best_loss > eval_res['loss']:
best_loss = eval_res['loss']
save_path = os.path.join(args.save_dir, 'model_best_loss.bin')
torch.save(model.state_dict(), save_path)
print("Best Loss : ", best_loss)
print("\n")
if last_update is None or best_acc < eval_res['joint_acc']:
best_acc = eval_res['joint_acc']
save_path = os.path.join(args.save_dir, 'model_best_acc.bin')
torch.save(model.state_dict(), save_path)
last_update = epoch
print("Best Acc : ", best_acc)
print("\n")
logger.info(
"*** Epoch=%d, Last Update=%d, Dev Loss=%.6f, Dev Acc=%.6f, Dev Turn Acc=%.6f, Best Loss=%.6f, Best Acc=%.6f ***"
% (epoch, last_update, eval_res['loss'], eval_res['joint_acc'],
eval_res['joint_turn_acc'], best_loss, best_acc))
if (epoch + 1) % args.eval_epoch == 0:
eval_res = model_evaluation(model, test_data_raw, tokenizer,
slot_meta, label_list, epoch + 1)
logger.info(
"*** Epoch=%d, Last Update=%d, Tes Loss=%.6f, Tes Acc=%.6f, Tes Turn Acc=%.6f, Best Loss=%.6f, Best Acc=%.6f ***"
% (epoch, last_update, eval_res['loss'], eval_res['joint_acc'],
eval_res['joint_turn_acc'], best_loss, best_acc))
if last_update + args.patience <= epoch:
break
print("Test using best loss model...")
best_epoch = 0
ckpt_path = os.path.join(args.save_dir, 'model_best_loss.bin')
model = BeliefTracker(args, slot_lookup, value_lookup, num_labels,
slot_value_pos, device)
ckpt = torch.load(ckpt_path, map_location='cpu')
model.load_state_dict(ckpt)
model.to(device)
test_res = model_evaluation(model,
test_data_raw,
tokenizer,
slot_meta,
label_list,
best_epoch,
is_gt_p_state=False)
logger.info("Results based on best loss: ")
logger.info(test_res)
#----------------------------------------------------------------------
print("Test using best acc model...")
ckpt_path = os.path.join(args.save_dir, 'model_best_acc.bin')
model = BeliefTracker(args, slot_lookup, value_lookup, num_labels,
slot_value_pos, device)
ckpt = torch.load(ckpt_path, map_location='cpu')
model.load_state_dict(ckpt)
model.to(device)
test_res = model_evaluation(model,
test_data_raw,
tokenizer,
slot_meta,
label_list,
best_epoch + 1,
is_gt_p_state=False)
logger.info("Results based on best acc: ")
logger.info(test_res)