一起读Bert文本分类代码(pytorch篇六）

Bert是去年google发布的新模型，打破了11项纪录，关于模型基础部分就不在这篇文章里多说了。这次想和大家一起读的是huggingface的pytorch-pretrained-BERT代码examples里的文本分类任务run_classifier。

关于源代码可以在huggingface的github中找到。

huggingface/pytorch-pretrained-BERT​ github.com

在前五篇文章中我分别介绍了数据预处理部分和模型部分：

周剑：一起读Bert文本分类代码 (pytorch篇 一）​ zhuanlan.zhihu.com
周剑：一起读Bert文本分类代码 (pytorch篇 二）​ zhuanlan.zhihu.com
周剑：一起读Bert文本分类代码 (pytorch篇 三）​ zhuanlan.zhihu.com
周剑：一起读Bert文本分类代码 (pytorch篇 四）​ zhuanlan.zhihu.com
周剑：一起读Bert文本分类代码 (pytorch篇 五）​ zhuanlan.zhihu.com

在这篇文章中，我会带大家继续读examples.run_classifier.py文件中的训练以及预测部分。

我们接着model继续往下读：

if args.fp16:
model.half()
model.to(device)
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())
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}
]
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=num_train_optimization_steps)

这段代码主要是处理单精度和双精度，把模型添加到cpu或者gpu中，定义优化函数的作用。

其中可以选择的优化函数有FusedAdam，它是NVIDIA开源面向精简混合精度和分布式训练的Pytorch扩展的优化函数，具体可以去NVIDIA的apex包下查看。还可以选择BertAdam，它是自定义Adam优化器，具体的代码可以在pytorch_pretrained_bert.optimization.py中找到。

继续往下看主函数，这段代码主要是训练和保存模型：

global_step = 0
nb_tr_steps = 0
tr_loss = 0
if args.do_train:
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_optimization_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()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 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
loss = model(input_ids, segment_ids, input_mask, label_ids)
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:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(global_step/num_train_optimization_steps, args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
# 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")
if args.do_train:
torch.save(model_to_save.state_dict(), output_model_file)

这段代码的主要内容是执行前面定义数据与处理函数，将数据处理好后feed入模型中，自动计算backward并更新模型参数。

这里值得一提的地方有两个：第一，DistributedSampler是pytorch中将数据自动分发给多gpu的函数。pytorch在1.0版本后对多gpu训练进行了优化，现在已经很不错了。第二，在这段代码最后保存模型时，建议只保存模型参数，即模型本身。这样每次读取的时候就只需要读取模型参数，不需要从保存的文件中读取模型结构，能加快模型读取速度。但只保存模型参数的话，就需要自己先定义模型，然后读取参数。

继续读主函数，这段代码主要是预测用：

model_state_dict = torch.load(output_model_file)
model = BertForSequenceClassification.from_pretrained(args.bert_model, state_dict=model_state_dict, num_labels=num_labels)
model.to(device)
if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
eval_examples = processor.get_dev_examples(args.data_dir)
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
for input_ids, input_mask, segment_ids, label_ids in tqdm(eval_dataloader, desc="Evaluating"):
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():
tmp_eval_loss = model(input_ids, segment_ids, input_mask, label_ids)
logits = model(input_ids, segment_ids, input_mask)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
tmp_eval_accuracy = 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
eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
loss = tr_loss/nb_tr_steps if args.do_train else None
result = {'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'global_step': global_step,
'loss': loss}
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))

这段代码主要是读取模型并预测输出结果用。如果使用多gpu，还是会把数据计算自动分配到各个gpu上。但是，numpy()只能使用于cpu。所以，这里需要注意，如果使用gpu训练或者预测，之使用numpy记得转到cpu中。这段代码最后输出一个txt的预测文件。

这样就带大家读完了所有examples.run_classifier.py中的代码和其调用的Bert模型的代码。