def __init__(self, model_archived_file: str, cuda_device: str = "cpu"):
"""
model_archived_file: ends with "tar.gz"
OR
directly use the model folder patth
"""
device = torch.device(cuda_device)
if model_archived_file.endswith("tar.gz"):
tar = tarfile.open(model_archived_file)
self.conf = pickle.load(tar.extractfile(
tar.getnames()[1])) ## config file
self.model = TransformersCRF(self.conf)
self.model.load_state_dict(
torch.load(tar.extractfile(tar.getnames()[2]),
map_location=device)) ## model file
else:
folder_name = model_archived_file
assert os.path.isdir(folder_name)
f = open(folder_name + "/config.conf", 'rb')
self.conf = pickle.load(f)
f.close()
self.model = TransformersCRF(self.conf)
self.model.load_state_dict(
torch.load(f"{folder_name}/lstm_crf.m", map_location=device))
self.conf.device = device
self.model.to(device)
self.model.eval()
print(
colored(
f"[Data Info] Tokenizing the instances using '{self.conf.embedder_type}' tokenizer",
"blue"))
self.tokenizer = context_models[
self.conf.embedder_type]["tokenizer"].from_pretrained(
self.conf.embedder_type)
def evaluate_model(config: Config, model: TransformersCRF, data_loader: DataLoader, name: str, insts: List, print_each_type_metric: bool = False):
## evaluation
p_dict, total_predict_dict, total_entity_dict = Counter(), Counter(), Counter()
batch_size = data_loader.batch_size
with torch.no_grad():
for batch_id, batch in tqdm(enumerate(data_loader, 0), desc="--evaluating batch", total=len(data_loader)):
one_batch_insts = insts[batch_id * batch_size:(batch_id + 1) * batch_size]
batch_max_scores, batch_max_ids = model.decode(words= batch.input_ids.to(config.device),
word_seq_lens = batch.word_seq_len.to(config.device),
orig_to_tok_index = batch.orig_to_tok_index.to(config.device),
input_mask = batch.attention_mask.to(config.device))
batch_p , batch_predict, batch_total = evaluate_batch_insts(one_batch_insts, batch_max_ids, batch.label_ids, batch.word_seq_len, config.idx2labels)
p_dict += batch_p
total_predict_dict += batch_predict
total_entity_dict += batch_total
batch_id += 1
f1Scores = []
if print_each_type_metric or config.print_detail_f1 or (config.earlystop_atr == "macro"):
for key in total_entity_dict:
precision_key, recall_key, fscore_key = get_metric(p_dict[key], total_entity_dict[key], total_predict_dict[key])
print(f"[{key}] Prec.: {precision_key:.2f}, Rec.: {recall_key:.2f}, F1: {fscore_key:.2f}")
f1Scores.append(fscore_key)
if len(f1Scores) > 0:
print(f"[{name} set Total] Macro F1: {sum(f1Scores) / len(f1Scores):.2f}")
total_p = sum(list(p_dict.values()))
total_predict = sum(list(total_predict_dict.values()))
total_entity = sum(list(total_entity_dict.values()))
precision, recall, fscore = get_metric(total_p, total_entity, total_predict)
print(colored(f"[{name} set Total] Prec.: {precision:.2f}, Rec.: {recall:.2f}, Micro F1: {fscore:.2f}", 'blue'), flush=True)
if config.earlystop_atr == "macro" and len(f1Scores) > 0:
fscore = sum(f1Scores) / len(f1Scores)
return [precision, recall, fscore]
class TransformersNERPredictor:
def __init__(self, model_archived_file:str,
cuda_device: str = "cpu"):
"""
model_archived_file: ends with "tar.gz"
OR
directly use the model folder patth
"""
device = torch.device(cuda_device)
if model_archived_file.endswith("tar.gz"):
tar = tarfile.open(model_archived_file)
self.conf = pickle.load(tar.extractfile(tar.getnames()[1])) ## config file
self.model = TransformersCRF(self.conf)
self.model.load_state_dict(torch.load(tar.extractfile(tar.getnames()[2]), map_location=device)) ## model file
else:
folder_name = model_archived_file
assert os.path.isdir(folder_name)
f = open(folder_name + "/config.conf", 'rb')
self.conf = pickle.load(f)
f.close()
self.model = TransformersCRF(self.conf)
self.model.load_state_dict(torch.load(f"{folder_name}/lstm_crf.m", map_location=device))
self.conf.device = device
self.model.to(device)
self.model.eval()
print(colored(f"[Data Info] Tokenizing the instances using '{self.conf.embedder_type}' tokenizer", "blue"))
self.tokenizer = context_models[self.conf.embedder_type]["tokenizer"].from_pretrained(self.conf.embedder_type)
def predict(self, sents: List[List[str]], batch_size = -1):
batch_size = len(sents) if batch_size == -1 else batch_size
dataset = TransformersNERDataset(file=None, sents=sents, tokenizer=self.tokenizer, label2idx=self.conf.label2idx, is_train=False)
loader = DataLoader(dataset, batch_size=batch_size, shuffle=False, num_workers=1, collate_fn=dataset.collate_fn)
all_predictions = []
for batch_id, batch in tqdm(enumerate(loader, 0), desc="--evaluating batch", total=len(loader)):
one_batch_insts = dataset.insts[batch_id * batch_size:(batch_id + 1) * batch_size]
batch_max_scores, batch_max_ids = self.model.decode(words= batch.input_ids.to(self.conf.device),
word_seq_lens = batch.word_seq_len.to(self.conf.device),
orig_to_tok_index = batch.orig_to_tok_index.to(self.conf.device),
input_mask = batch.attention_mask.to(self.conf.device))
for idx in range(len(batch_max_ids)):
length = batch.word_seq_len[idx]
prediction = batch_max_ids[idx][:length].tolist()
prediction = prediction[::-1]
prediction = [self.conf.idx2labels[l] for l in prediction]
one_batch_insts[idx].prediction = prediction
all_predictions.append(prediction)
return all_predictions
def main():
parser = argparse.ArgumentParser(description="LSTM CRF implementation")
opt = parse_arguments(parser)
if opt.mode == "train":
conf = Config(opt)
set_seed(opt, conf.seed)
print(colored(f"[Data Info] Tokenizing the instances using '{conf.embedder_type}' tokenizer", "blue"))
tokenizer = context_models[conf.embedder_type]["tokenizer"].from_pretrained(conf.embedder_type, add_prefix_space=True)
print(colored(f"[Data Info] Reading dataset from: \n{conf.train_file}\n{conf.dev_file}\n{conf.test_file}", "blue"))
train_dataset = TransformersNERDataset(conf.train_file, tokenizer, number=conf.train_num, is_train=True)
conf.label2idx = train_dataset.label2idx
conf.idx2labels = train_dataset.idx2labels
dev_dataset = TransformersNERDataset(conf.dev_file, tokenizer, number=conf.dev_num, label2idx=train_dataset.label2idx, is_train=False)
test_dataset = TransformersNERDataset(conf.test_file, tokenizer, number=conf.test_num, label2idx=train_dataset.label2idx, is_train=False)
num_workers = 8
conf.label_size = len(train_dataset.label2idx)
train_dataloader = DataLoader(train_dataset, batch_size=conf.batch_size, shuffle=True, num_workers=num_workers,
collate_fn=train_dataset.collate_fn)
dev_dataloader = DataLoader(dev_dataset, batch_size=conf.batch_size, shuffle=False, num_workers=num_workers,
collate_fn=dev_dataset.collate_fn)
test_dataloader = DataLoader(test_dataset, batch_size=conf.batch_size, shuffle=False, num_workers=num_workers,
collate_fn=test_dataset.collate_fn)
train_model(conf, conf.num_epochs, train_dataloader, dev_dataloader, test_dataloader)
else:
folder_name = f"model_files/{opt.model_folder}"
device = torch.device(opt.device)
assert os.path.isdir(folder_name)
f = open(folder_name + "/config.conf", 'rb')
saved_config = pickle.load(f) # we use `label2idx` from old config, but test file, test number
f.close()
print(colored(f"[Data Info] Tokenizing the instances using '{saved_config.embedder_type}' tokenizer", "blue"))
tokenizer = context_models[saved_config.embedder_type]["tokenizer"].from_pretrained(saved_config.embedder_type, add_prefix_space=True)
test_dataset = TransformersNERDataset(opt.test_file, tokenizer, number=opt.test_num,
label2idx=saved_config.label2idx, is_train=False)
test_dataloader = DataLoader(test_dataset, batch_size=opt.batch_size, shuffle=False, num_workers=1,
collate_fn=test_dataset.collate_fn)
model = TransformersCRF(saved_config)
model.load_state_dict(torch.load(f"{folder_name}/lstm_crf.m", map_location=device))
model.eval()
evaluate_model(config=saved_config, model=model, data_loader=test_dataloader, name="test mode", insts = test_dataset.insts,
print_each_type_metric=False)
def train_model(config: Config, epoch: int, train_loader: DataLoader, dev_loader: DataLoader, test_loader: DataLoader):
### Data Processing Info
train_num = len(train_loader)
print(f"[Data Info] number of training instances: {train_num}")
print(
colored(f"[Model Info]: Working with transformers package from huggingface with {config.embedder_type}", 'red'))
print(colored(f"[Optimizer Info]: You should be aware that you are using the optimizer from huggingface.", 'red'))
print(colored(f"[Optimizer Info]: Change the optimier in transformers_util.py if you want to make some modifications.", 'red'))
model = TransformersCRF(config)
optimizer, scheduler = get_huggingface_optimizer_and_scheduler(config, model, num_training_steps=len(train_loader) * epoch,
weight_decay=0.0, eps = 1e-8, warmup_step=0)
print(colored(f"[Optimizer Info] Modify the optimizer info as you need.", 'red'))
print(optimizer)
model.to(config.device)
best_dev = [-1, 0]
best_test = [-1, 0]
model_folder = config.model_folder
res_folder = "results"
if os.path.exists("model_files/" + model_folder):
raise FileExistsError(
f"The folder model_files/{model_folder} exists. Please either delete it or create a new one "
f"to avoid override.")
model_path = f"model_files/{model_folder}/lstm_crf.m"
config_path = f"model_files/{model_folder}/config.conf"
res_path = f"{res_folder}/{model_folder}.results"
print("[Info] The model will be saved to: %s.tar.gz" % (model_folder))
os.makedirs(f"model_files/{model_folder}", exist_ok= True) ## create model files. not raise error if exist
os.makedirs(res_folder, exist_ok=True)
no_incre_dev = 0
print(colored(f"[Train Info] Start training, you have set to stop if performace not increase for {config.max_no_incre} epochs",'red'))
for i in tqdm(range(1, epoch + 1), desc="Epoch"):
epoch_loss = 0
start_time = time.time()
model.zero_grad()
model.train()
for iter, batch in tqdm(enumerate(train_loader, 1), desc="--training batch", total=len(train_loader)):
optimizer.zero_grad()
loss = model(words = batch.input_ids.to(config.device), word_seq_lens = batch.word_seq_len.to(config.device),
orig_to_tok_index = batch.orig_to_tok_index.to(config.device), input_mask = batch.attention_mask.to(config.device),
labels = batch.label_ids.to(config.device))
epoch_loss += loss.item()
loss.backward()
if config.max_grad_norm > 0:
torch.nn.utils.clip_grad_norm_(model.parameters(), config.max_grad_norm)
optimizer.step()
optimizer.zero_grad()
scheduler.step()
model.zero_grad()
end_time = time.time()
print("Epoch %d: %.5f, Time is %.2fs" % (i, epoch_loss, end_time - start_time), flush=True)
model.eval()
dev_metrics = evaluate_model(config, model, dev_loader, "dev", dev_loader.dataset.insts)
test_metrics = evaluate_model(config, model, test_loader, "test", test_loader.dataset.insts)
if dev_metrics[2] > best_dev[0]:
print("saving the best model...")
no_incre_dev = 0
best_dev[0] = dev_metrics[2]
best_dev[1] = i
best_test[0] = test_metrics[2]
best_test[1] = i
torch.save(model.state_dict(), model_path)
# Save the corresponding config as well.
f = open(config_path, 'wb')
pickle.dump(config, f)
f.close()
write_results(res_path, test_loader.dataset.insts)
else:
no_incre_dev += 1
model.zero_grad()
if no_incre_dev >= config.max_no_incre:
print("early stop because there are %d epochs not increasing f1 on dev"%no_incre_dev)
break
print("Archiving the best Model...")
with tarfile.open(f"model_files/{model_folder}.tar.gz", "w:gz") as tar:
tar.add(f"model_files/{model_folder}", arcname=os.path.basename(model_folder))
print("Finished archiving the models")
print("The best dev: %.2f" % (best_dev[0]))
print("The corresponding test: %.2f" % (best_test[0]))
print("Final testing.")
model.load_state_dict(torch.load(model_path))
model.eval()
evaluate_model(config, model, test_loader, "test", test_loader.dataset.insts)
write_results(res_path, test_loader.dataset.insts)
def evaluate_model(config: Config,
model: TransformersCRF,
data_loader: DataLoader,
name: str,
insts: List,
print_each_type_metric: bool = False):
## evaluation
#p_dict, total_predict_dict, total_entity_dict = Counter(), Counter(), Counter()
f1_metrics = F1Measure()
batch_size = data_loader.batch_size
with torch.no_grad():
with tqdm(enumerate(data_loader, 0),
desc="--evaluating batch",
total=len(data_loader)) as teval:
for batch_id, batch in teval:
one_batch_insts = insts[batch_id * batch_size:(batch_id + 1) *
batch_size]
batch_max_scores, batch_max_ids = model.decode(
words=batch.input_ids.to(config.device),
word_seq_lens=batch.word_seq_len.to(config.device),
orig_to_tok_index=batch.orig_to_tok_index.to(
config.device),
input_mask=batch.attention_mask.to(config.device))
batch_p, batch_predict, batch_total = evaluate_batch_insts(
one_batch_insts, batch_max_ids, batch.label_ids,
batch.word_seq_len, config.idx2labels)
#p_dict += batch_p
#total_predict_dict += batch_predict
#total_entity_dict += batch_total
f1_metrics.update(batch_p, batch_predict, batch_total)
teval.set_postfix(**f1_metrics.get_metric(
print_each_type_metric=False)[0])
batch_id += 1
final_metrics, final_metrics_key = f1_metrics.get_metric(
print_each_type_metric)
'''
if print_each_type_metric:
for key in total_entity_dict:
precision_key, recall_key, fscore_key = get_metric(p_dict[key], total_entity_dict[key], total_predict_dict[key])
print(f"[{key}] Prec.: {precision_key:.2f}, Rec.: {recall_key:.2f}, F1: {fscore_key:.2f}")
'''
if final_metrics_key is not None:
for key in final_metrics_key:
precision_key, recall_key, fscore_key = final_metrics_key[key][
"Prec."], final_metrics_key[key]["Recl."], final_metrics_key[
key]["F1"]
print(
f"[{key}] Prec.: {precision_key:.2f}, Rec.: {recall_key:.2f}, F1: {fscore_key:.2f}"
)
#total_p = sum(list(p_dict.values()))
#total_predict = sum(list(total_predict_dict.values()))
#total_entity = sum(list(total_entity_dict.values()))
#precision, recall, fscore = get_metric(total_p, total_entity, total_predict)
precision, recall, fscore = final_metrics["Prec"], final_metrics[
'Recl'], final_metrics["F1"]
print(colored(
f"[{name} set Total] Prec.: {precision:.2f}, Rec.: {recall:.2f}, F1: {fscore:.2f}",
'blue'),
flush=True)
return [precision, recall, fscore]