def load_and_cache_examples(args,
tokenizer,
processor,
label_list,
mode="train"):
if args.local_rank not in [-1, 0]:
# Make sure only the first process in distributed training process the dataset,
# and the others will use the cache
torch.distributed.barrier()
examples = processor.get_examples(args.data_dir, mode)
features = convert_examples_to_features(
examples,
label_list,
args.max_seq_length,
tokenizer,
args.output_mode,
cls_token_at_end=bool(args.model_type in ['xlnet']),
# xlnet has a cls token at the end
cls_token=tokenizer.cls_token,
cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0,
sep_token=tokenizer.sep_token,
sep_token_extra=bool(args.model_type in ['roberta']),
# roberta uses an extra separator b/w pairs of sentences,
# cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
pad_on_left=bool(args.model_type in ['xlnet']),
# pad on the left for xlnet
pad_token=tokenizer.encoder[tokenizer.pad_token] if args.model_type
in ['roberta'] else tokenizer.vocab[tokenizer.pad_token],
pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0,
)
if args.local_rank == 0:
# Make sure only the first process in distributed training process the dataset,
# and the others will use the cache
torch.distributed.barrier()
# Convert to Tensors and build dataset
all_input_ids = torch.tensor([f.input_ids for f in features],
dtype=torch.long)
all_entity1 = torch.tensor([f.e1_pos for f in features], dtype=torch.long)
all_entity2 = torch.tensor([f.e2_pos for f in features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in features],
dtype=torch.long)
if args.output_mode == "classification":
all_label_ids = torch.tensor([f.label_id for f in features],
dtype=torch.long)
elif args.output_mode == "regression":
all_label_ids = torch.tensor([f.label_id for f in features],
dtype=torch.float)
dataset = TensorDataset(all_input_ids, all_entity1, all_entity2,
all_input_mask, all_segment_ids, all_label_ids)
return dataset
def load_and_cache_examples(mode, train_batch_size, eval_batch_size):
'''
:param mode: "train","dev"
:return:
'''
"""构造迭代器"""
processor, tokenizer = init_params()
if args.local_rank not in [-1, 0]:
torch.distributed.barrier(
) # Make sure only the first process in distributed training process the dataset, and the others will use the cache
if mode == "train":
examples = processor.get_train_examples(args.data_dir)
#t_total
#num_train_steps = int(
# len(examples) / train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs)
batch_size = train_batch_size
#logger.info(" Num steps = %d", num_train_steps)
elif mode == 'dev':
examples = processor.get_dev_examples(args.data_dir)
batch_size = eval_batch_size
else:
raise ValueError("Invalid mode %s" % mode)
#label_list = processor.get_labels()
output_mode = "classification"
#特征
try:
if mode == "train":
with open(os.path.join(args.data_dir, args.TRAIN_US_FEATURE_FILE),
'rb') as f: #TRAIN_FEATURE_FILE
features = pickle.load(f)
else:
with open(os.path.join(args.data_dir, args.DEV_US_FEATURE_FILE),
'rb') as f: #DEV_FEATURE_FILE
features = pickle.load(f)
except:
features = convert_examples_to_features(
examples,
args.max_seq_length,
args.split_num,
tokenizer,
mode=mode,
cls_token_at_end=bool(args.model_type in ['xlnet']
), # xlnet has a cls token at the end
cls_token=tokenizer.cls_token,
sep_token=tokenizer.sep_token,
cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0,
pad_on_left=bool(
args.model_type in ['xlnet']), # pad on the left for xlnet
pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0)
logger.info(" Num examples = %d", len(examples))
logger.info(" Batch size = %d", batch_size)
# Convert to Tensors and build dataset
all_input_ids = torch.tensor(select_field(features, 'input_ids'),
dtype=torch.long)
all_input_mask = torch.tensor(select_field(features, 'input_mask'),
dtype=torch.long)
all_segment_ids = torch.tensor(select_field(features, 'segment_ids'),
dtype=torch.long)
#all_label = torch.tensor([f.label for f in features], dtype=torch.long)
# all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
# all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
# all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
if output_mode == "classification":
all_labels = torch.tensor([int(f.label) for f in features],
dtype=torch.long)
elif output_mode == "regression":
all_labels = torch.tensor([int(f.label) for f in features],
dtype=torch.float)
#数据集
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
all_labels)
if mode == "train":
sampler = RandomSampler(dataset) #作用近似shuffle
elif mode == 'dev':
sampler = SequentialSampler(dataset)
else:
raise ValueError("Invalid mode %s" % mode)
# 迭代器
iterator = DataLoader(dataset, sampler=sampler, batch_size=batch_size)
return iterator
def train(global_step, epoch):
train_features = convert_examples_to_features(train_examples,
label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
total_loss, total_accuracy = 0, 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
# 加上 label_ids 输出结果是 loss
# loss = model(input_ids, segment_ids, input_mask, label_ids)
out, loss = model(input_ids, segment_ids, input_mask,
label_ids) # [batch, 1]
logits = (out > args.threshold).type(torch.LongTensor)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
tmp_train_accuracy = compute_accuracy(logits, label_ids)
# print(tmp_train_accuracy)
total_loss += loss.mean().item()
total_accuracy += tmp_train_accuracy
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
# tr_loss += loss.item()
# nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
# modify learning rate with special warm up BERT uses
lr_this_step = args.learning_rate * warmup_linear(
global_step / (t_total), args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
args.log_interval = 300
if (step + 1) % args.log_interval == 0:
# print(out)
# print(pred)
# print(label)
logger.info("|----epoch {}, eclipse {}/{}, lr {:.4f},"
"loss {:.4f}, acc {:.4f}".format(
epoch, step + 1, len(train_dataloader),
lr_this_step, total_loss / args.log_interval,
total_accuracy / args.log_interval))
total_loss, total_accuracy = 0.0, 0.0
def main(debug=True):
# device
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
# distribution training
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
# gradient accumulation
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
# use the same random seed
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
# output dir is need
#if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
# raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
#os.makedirs(args.output_dir, exist_ok=True)
# data processor
processor = QuoraProcessor()
# processor = MrpcProcessor()
num_labels = 2
label_list = processor.get_labels()
# BertTokenizer
# tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
tokenizer = BertTokenizer.from_pretrained(
"./pre_trained_models/bert-base-uncased-vocab.txt")
train_examples = None
num_train_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir,
debug=debug,
debug_length=100)
num_train_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
print(len(train_examples))
model = Bce_model()
if args.fp16:
model.half()
model.to(device)
# distributed parallel
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
# print(param_optimizer)
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
}]
t_total = num_train_steps
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=t_total)
def train(global_step, epoch):
train_features = convert_examples_to_features(train_examples,
label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
total_loss, total_accuracy = 0, 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
# 加上 label_ids 输出结果是 loss
# loss = model(input_ids, segment_ids, input_mask, label_ids)
out, loss = model(input_ids, segment_ids, input_mask,
label_ids) # [batch, 1]
logits = (out > args.threshold).type(torch.LongTensor)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
tmp_train_accuracy = compute_accuracy(logits, label_ids)
# print(tmp_train_accuracy)
total_loss += loss.mean().item()
total_accuracy += tmp_train_accuracy
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
# tr_loss += loss.item()
# nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
# modify learning rate with special warm up BERT uses
lr_this_step = args.learning_rate * warmup_linear(
global_step / (t_total), args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
args.log_interval = 300
if (step + 1) % args.log_interval == 0:
# print(out)
# print(pred)
# print(label)
logger.info("|----epoch {}, eclipse {}/{}, lr {:.4f},"
"loss {:.4f}, acc {:.4f}".format(
epoch, step + 1, len(train_dataloader),
lr_this_step, total_loss / args.log_interval,
total_accuracy / args.log_interval))
total_loss, total_accuracy = 0.0, 0.0
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
# get dev examples anf dev features
eval_examples = processor.get_dev_examples(args.data_dir,
debug=debug,
debug_length=16)
eval_features = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
def eval():
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
out_all = []
labels_all = []
for i, (input_ids, input_mask, segment_ids,
label_ids) in enumerate(eval_dataloader):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
out, tmp_eval_loss = model(input_ids, segment_ids, input_mask,
label_ids)
eval_loss += tmp_eval_loss.mean().item()
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
# print(out.shape)
out_all.append(out)
labels_all.append(label_ids)
eval_loss = eval_loss / nb_eval_steps
model.train()
return eval_loss, out_all, labels_all
if args.do_train and args.do_eval:
for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
global_step = 0
train(global_step, epoch)
best_acc, best_f1 = 0.0, 0.0
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_loss, out_all, labels_all = eval()
out_all = torch.cat(out_all, dim=0)
labels_all = torch.cat(labels_all, dim=0)
labels_eval = labels_all.cpu().numpy()
for threshold in np.linspace(0.2, 0.6, 40):
logits_eval = (out_all > threshold).type(torch.LongTensor)
logits_eval = logits_eval.detach().cpu().numpy()
eval_accuracy = compute_accuracy(logits_eval, labels_eval)
eval_f1, eval_precision, eval_recall = compute_f1_precision_recall(
logits_eval, labels_eval)
if eval_f1 > best_f1:
best_f1 = eval_f1
# Save a trained model
model_to_save = model.module if hasattr(
model,
'module') else model # Only save the model it-self
output_model_file = os.path.join(
args.output_dir, "pytorch_model.bin")
torch.save(model_to_save.state_dict(),
output_model_file)
logger.info(
'epcoh {:d}, threshold {:.4f}, accuracy {:.4f}, precision {:.4f}, recall {:.4f}, f1 {:.4f}, best_f1 {:.4f}'
.format(epoch, threshold, eval_accuracy,
eval_precision, eval_recall, eval_f1, best_f1))
if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank()
== 0) and not args.do_train:
# Load a trained model that you have fine-tuned
model_state_dict = torch.load(output_model_file)
model = BertForSequenceClassification.from_pretrained(
args.bert_model, state_dict=model_state_dict)
model.to(device)
eval_loss, out_all, labels_all = eval()
def eval():
eval_examples = processor.get_dev_examples(args.data_dir,
debug=debug,
debug_length=100)
eval_features = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
logits_all = []
labels_all = []
for i, (input_ids, input_mask, segment_ids,
label_ids) in enumerate(eval_dataloader):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
# 加上 label_ids 输出是 loss, 不加 label_ids 输出是 logits, [batch, 2]
tmp_eval_loss = model(input_ids, segment_ids, input_mask,
label_ids)
logits = model(input_ids, segment_ids, input_mask)
# print(logits)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
tmp_eval_accuracy = compute_accuracy(logits, label_ids)
eval_loss += tmp_eval_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
logits_all.append(logits)
labels_all.append(label_ids)
eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
# return eval_loss, eval_accuracy
logits_eval = np.concatenate(logits_all, axis=0)
labels_eval = np.concatenate(labels_all, axis=0)
# print(logits_eval)
# print(labels_eval)
eval_f1, eval_precision, eval_recall = compute_f1_precision_recall(
logits_eval, labels_eval)
model.train()
return eval_loss, eval_accuracy, eval_f1, eval_precision, eval_recall
def preprocess():
'''
针对测试集数据预测
:return:
'''
model = load_model(os.path.join(args.ROOT_DIR, args.output_dir),
args.model_type)
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend='nccl')
n_gpu = 1
model.to(device)
processors = {"sentiment_analysis": SentiAnalysisProcessor}
task_name = args.task_name.lower()
processor = processors[task_name]()
examples = processor.get_test_examples(args.data_dir)
tokenizer = XLNetTokenizer.from_pretrained(
os.path.join(args.ROOT_DIR, args.xlnet_model),
do_lower_case=args.do_lower_case)
mode = 'test'
try:
if mode == 'test':
with open(os.path.join(args.data_dir, args.TEST_FEATURE_FILE),
'rb') as f: # TRAIN_FEATURE_FILE
features = pickle.load(f)
except:
features = convert_examples_to_features(
examples,
args.max_seq_length,
args.split_num,
tokenizer,
mode=mode,
cls_token_at_end=bool(args.model_type in ['xlnet']),
# xlnet has a cls token at the end
cls_token=tokenizer.cls_token,
sep_token=tokenizer.sep_token,
cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0,
pad_on_left=bool(
args.model_type in ['xlnet']), # pad on the left for xlnet
pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0)
all_input_ids = torch.tensor(select_field(features, 'input_ids'),
dtype=torch.long)
all_input_mask = torch.tensor(select_field(features, 'input_mask'),
dtype=torch.long)
all_segment_ids = torch.tensor(select_field(features, 'segment_ids'),
dtype=torch.long)
#数据集
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids)
sampler = SequentialSampler(dataset) #作用近似shuffle
dataloader = DataLoader(dataset,
sampler=sampler,
batch_size=args.per_gpu_train_batch_size *
max(1, n_gpu))
model.eval()
y_predicts = []
for input_ids, input_mask, segment_ids in dataloader:
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
with torch.no_grad():
logits = model(input_ids=input_ids,
token_type_ids=segment_ids,
attention_mask=input_mask)[0]
predicts = model.predict(logits)
y_predicts.append(torch.from_numpy(predicts))
eval_predicted = torch.cat(y_predicts, dim=0).cpu().numpy()
df = pd.read_csv(os.path.join(args.data_dir, args.TEST_CORPUS_FILE))
df['labels'] = eval_predicted
df[['id', 'labels']].to_csv('./data/test_final.csv',
sep=',',
encoding='utf_8_sig',
header=True,
index=False)