Transformers解决文本分类任务、超参搜索

本文主要内容转自天国之影笔记Task06，之后具体的API进行了一些查询，写了一些说明。
@[toc]

1 文本分类任务简介

• 使用Transformers代码库中的模型来解决文本分类任务，任务来源于GLUE Benchmark
• GLUE榜单的9个级别的分类任务：
  1. CoLA (Corpus of Linguistic Acceptability)：鉴别一个句子是否语法正确.
  2. MNLI (Multi-Genre Natural Language Inference)：给定一个假设，判断另一个句子与该假设的关系：entails、contradicts、unrelated。
  3. MRPC (Microsoft Research Paraphrase Corpus)：判断两个句子是否互为paraphrases
  4. QNLI (Question-answering Natural Language Inference)：判断第2句是否包含第1句问题的答案
  5. QQP (Quora Question Pairs2)：判断两个问句是否语义相同
  6. RTE (Recognizing Textual Entailment)：判断一个句子是否与假设成entail关系
  7. SST-2 (Stanford Sentiment Treebank)：判断一个句子的情感正负向
  8. STS-B (Semantic Textual Similarity Benchmark)：判断两个句子的相似性（分数为1-5分）
  9. WNLI (Winograd Natural Language Inference)：判断带有匿名代词的句子中，是否存在能够替换该代词的子句

     GLUE_TASKS = ["cola", "mnli", "mnli-mm", "mrpc",
                   "qnli", "qqp", "rte", "sst2", "stsb", "wnli"]

     # 任务为CoLA任务
     task = "cola"
     # BERT模型
     model_checkpoint = "distilbert-base-uncased"
     # 根据GPU调整batch_size大小，避免显存溢出
     batch_size = 16

     2 加载数据

     2.1 加载数据和对应的评测方式

     #数据加载和评测方式加载只需要简单使用load_dataset和load_metric即可
     from datasets import load_dataset, load_metric

     ##根据任务名称加载数据和评估方法
     #除了mnli-mm以外，其他任务都可以直接通过任务名字进行加载。数据加载之后会自动缓存。
     actual_task = "mnli" if task == "mnli-mm" else task
     # 加载GLUE数据集
     dataset = load_dataset("glue", actual_task)
     # 加载GLUE的评测方式
     metric = load_metric('glue', actual_task)

     Reusing dataset glue (C:\Users\hurui.cache\huggingface\datasets\glue\cola\1.0.0\dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad)

2.2 查看数据

# 对于训练集、验证集和测试集，只需要使用对应的key（train，validation，test）即可得到相应的数据
dataset

DatasetDict({
    train: Dataset({
        features: ['sentence', 'label', 'idx'],
        num_rows: 8551
    })
    validation: Dataset({
        features: ['sentence', 'label', 'idx'],
        num_rows: 1043
    })
    test: Dataset({
        features: ['sentence', 'label', 'idx'],
        num_rows: 1063
    })
})

# 查看训练集第一条数据
dataset["train"][0]

{'sentence': "Our friends won't buy this analysis, let alone the next one we propose.",
 'label': 1,
 'idx': 0}

为了能够进一步理解数据长什么样子，下面的函数将从数据集里随机选择几个例子进行展示。

import datasets
import random
import pandas as pd
from IPython.display import display, HTML

def show_random_elements(dataset, num_examples=10):
    """从数据集中随机选择几条数据"""
    assert num_examples <= len(
        dataset), "Can't pick more elements than there are in the dataset."
    picks = []
    for _ in range(num_examples):
        pick = random.randint(0, len(dataset)-1)
        while pick in picks:
            pick = random.randint(0, len(dataset)-1)
        picks.append(pick)

    df = pd.DataFrame(dataset[picks])
    for column, typ in dataset.features.items():
        if isinstance(typ, datasets.ClassLabel):
            df[column] = df[column].transform(lambda i: typ.names[i])
    display(HTML(df.to_html()))

show_random_elements(dataset["train"])

	sentence	label	idx
0	No one can forgive you that comment.	acceptable	2078
1	Bill and Kathy married.	acceptable	2318
2	$5 will buy a ticket.	acceptable	2410
3	Which books did Robin talk to Chris and read?	unacceptable	7039
4	Jill offered the ball towards Bob.	unacceptable	2053
5	Has not Henri studied for his exam?	unacceptable	7466
6	Fanny stopped talking when in came Aunt Norris.	unacceptable	6778
7	Who do you think that would be nominated for the position?	unacceptable	4784
8	Mickey teamed with the women up.	unacceptable	440
9	Baseballs toss easily.	unacceptable	2783

2.3 查看评测方法

评估metic是datasets.Metric的一个实例:

metric

    Metric(name: "glue", features: {'predictions': Value(dtype='int64', id=None), 'references': Value(dtype='int64', id=None)}, usage: """
    Compute GLUE evaluation metric associated to each GLUE dataset.
    Args:
        predictions: list of predictions to score.
            Each translation should be tokenized into a list of tokens.
        references: list of lists of references for each translation.
            Each reference should be tokenized into a list of tokens.
    Returns: depending on the GLUE subset, one or several of:
        "accuracy": Accuracy
        "f1": F1 score
        "pearson": Pearson Correlation
        "spearmanr": Spearman Correlation
        "matthews_correlation": Matthew Correlation
    Examples:
    
        >>> glue_metric = datasets.load_metric('glue', 'sst2')  # 'sst2' or any of ["mnli", "mnli_mismatched", "mnli_matched", "qnli", "rte", "wnli", "hans"]
        >>> references = [0, 1]
        >>> predictions = [0, 1]
        >>> results = glue_metric.compute(predictions=predictions, references=references)
        >>> print(results)
        {'accuracy': 1.0}
    
        >>> glue_metric = datasets.load_metric('glue', 'mrpc')  # 'mrpc' or 'qqp'
        >>> references = [0, 1]
        >>> predictions = [0, 1]
        >>> results = glue_metric.compute(predictions=predictions, references=references)
        >>> print(results)
        {'accuracy': 1.0, 'f1': 1.0}
    
        >>> glue_metric = datasets.load_metric('glue', 'stsb')
        >>> references = [0., 1., 2., 3., 4., 5.]
        >>> predictions = [0., 1., 2., 3., 4., 5.]
        >>> results = glue_metric.compute(predictions=predictions, references=references)
        >>> print({"pearson": round(results["pearson"], 2), "spearmanr": round(results["spearmanr"], 2)})
        {'pearson': 1.0, 'spearmanr': 1.0}
    
        >>> glue_metric = datasets.load_metric('glue', 'cola')
        >>> references = [0, 1]
        >>> predictions = [0, 1]
        >>> results = glue_metric.compute(predictions=predictions, references=references)
        >>> print(results)
        {'matthews_correlation': 1.0}
    """, stored examples: 0)

调用metric的compute方法，传入labels和predictions即可得到metric的值：

#这里只是一个示例
import numpy as np
fake_preds = np.random.randint(0, 2, size=(64,))
fake_labels = np.random.randint(0, 2, size=(64,))
metric.compute(predictions=fake_preds, references=fake_labels)

{‘matthews_correlation’: -0.00392156862745098}

2.4 文本分类任务与评测方法

任务	评测方法
CoLA	Matthews Correlation Coefficient
MNLI	Accuracy
MRPC	Accuracy and F1 score
QNLI	Accuracy
QQP	Accuracy and F1 score
RTE	Accuracy
SST-2	Accuracy
STS-B	Pearson Correlation Coefficient and Spearman’s_Rank_Correlation_Coefficient
WNLI	Accuracy

3 数据预处理

3.1 数据预处理流程

• 使用工具：Tokenizer
• 流程：
  1. 对输入数据进行tokenize，得到tokens
  2. 将tokens转化为预训练模型中需要对应的token ID
  3. 将token ID转化为模型需要的输入格式

  为了达到数据预处理的目的，我们使用AutoTokenizer.from_pretrained方法实例化我们的tokenizer，这样可以确保：

• 我们得到一个与预训练模型一一对应的tokenizer。
• 使用指定的模型checkpoint对应的tokenizer的时候，我们也下载了模型需要的词表库vocabulary，准确来说是tokens vocabulary。
• 这个被下载的tokens vocabulary会被缓存起来，从而再次使用的时候不会重新下载

  3.2 构建模型对应的tokenizer

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(model_checkpoint, use_fast=True)

AutoTokenizer是一个通用的分词器类，使用AutoTokenizer.from_pretrained类方法实例化具体的分词器之一。

AutoTokenizer.from_pretrained(
   pretrained_model_name_or_path,
   *inputs,
   **kwargs,)

  注意：use_fast=True要求tokenizer必须是transformers.PreTrainedTokenizerFast类型，因为我们在预处理的时候需要用到fast tokenizer的一些特殊特性（比如多线程快速tokenizer）。如果对应的模型没有fast tokenizer，去掉这个选项即可。

  tokenizer既可以对单个文本进行预处理，也可以对一对文本进行预处理，tokenizer预处理后得到的数据满足预训练模型输入格式。这取决于我们选择的预训练模型，我们将会看到tokenizer有不同的返回，==tokenizer和预训练模型是一一对应的==，更多信息可以在这里进行学习。

tokenizer("Hello, this one sentence!", "And this sentence goes with it.")

{'input_ids': [101, 7592, 1010, 2023, 2028, 6251, 999, 102, 1998, 2023, 6251, 3632, 2007, 2009, 1012, 102], 'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}

3.3 对数据集datasets所有样本进行预处理

为了预处理我们的数据，我们需要知道不同数据和对应的数据格式，因此我们定义下面这个dict

# 定义如下dict，用于对数据格式进行检查
task_to_keys = {
    "cola": ("sentence", None),
    "mnli": ("premise", "hypothesis"),
    "mnli-mm": ("premise", "hypothesis"),
    "mrpc": ("sentence1", "sentence2"),
    "qnli": ("question", "sentence"),
    "qqp": ("question1", "question2"),
    "rte": ("sentence1", "sentence2"),
    "sst2": ("sentence", None),
    "stsb": ("sentence1", "sentence2"),
    "wnli": ("sentence1", "sentence2"),
}

# 对训练数据集的第1条数据进行数据格式检查
sentence1_key, sentence2_key = task_to_keys[task]
if sentence2_key is None:
    print(f"Sentence: {dataset['train'][0][sentence1_key]}")
else:
    print(f"Sentence 1: {dataset['train'][0][sentence1_key]}")
    print(f"Sentence 2: {dataset['train'][0][sentence2_key]}")

Sentence: Our friends won't buy this analysis, let alone the next one we propose.

# 构造数据预处理函数
def preprocess_function(examples):
    if sentence2_key is None:
        return tokenizer(examples[sentence1_key], truncation=True)
    return tokenizer(examples[sentence1_key], examples[sentence2_key], truncation=True)

  预处理函数可以处理单个样本，也可以对多个样本进行处理。如果输入是多个样本，那么返回的是一个list。
  使用map函数将预处理函数应用到（map)所有样本上。

# 对所有数据进行预处理
encoded_dataset = dataset.map(preprocess_function, batched=True)

Loading cached processed dataset at C:\Users\hurui\.cache\huggingface\datasets\glue\cola\1.0.0\dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad\cache-fd5eee62c2b8c26e.arrow
Loading cached processed dataset at C:\Users\hurui\.cache\huggingface\datasets\glue\cola\1.0.0\dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad\cache-0ce499346cf9c20b.arrow

HBox(children=(FloatProgress(value=0.0, max=2.0), HTML(value='')))

4 微调预训练模型

  既然我们是做seq2seq任务，那么我们需要使用AutoModelForSequenceClassification 这个类。和tokenizer相似，from_pretrained方法同样可以帮助我们下载并加载模型，同时也会对模型进行缓存，就不会重复下载模型啦。

4.1 加载分类模型

#STS-B是一个回归问题，MNLI是一个3分类问题
from transformers import AutoModelForSequenceClassification, TrainingArguments, Trainer

num_labels = 3 if task.startswith("mnli") else 1 if task == "stsb" else 2
model = AutoModelForSequenceClassification.from_pretrained(
    model_checkpoint, num_labels=num_labels)
#模型是最前面设置的model_checkpoint=distilbert-base-uncased

Some weights of the model checkpoint at distilbert-base-uncased were not used when initializing DistilBertForSequenceClassification: ['vocab_layer_norm.weight', 'vocab_transform.weight', 'vocab_projector.bias', 'vocab_projector.weight', 'vocab_transform.bias', 'vocab_layer_norm.bias']
- This IS expected if you are initializing DistilBertForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
- This IS NOT expected if you are initializing DistilBertForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
Some weights of DistilBertForSequenceClassification were not initialized from the model checkpoint at distilbert-base-uncased and are newly initialized: ['pre_classifier.bias', 'classifier.weight', 'classifier.bias', 'pre_classifier.weight']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.

  由于我们微调的任务是文本分类任务，而我们加载的是预训练的语言模型，所以会提示我们加载模型的时候扔掉了一些不匹配的神经网络参数（比如：预训练语言模型的神经网络head被扔掉了，同时随机初始化了文本分类的神经网络head）

4.2 设定训练参数

  Trainer训练工具需要3个要素，最重要的是训练的设定/参数 TrainingArguments。这个训练设定包含了能够定义训练过程的所有属性。

metric_name = "pearson" if task == "stsb" else "matthews_correlation" if task == "cola" else "accuracy"

args = TrainingArguments(
    #args包含了能够定义训练过程的所有属性
metric_name = "pearson" if task == "stsb" else "matthews_correlation" if task == "cola" else "accuracy"
args = TrainingArguments(
    "test-glue",#输出路径
    evaluation_strategy = "epoch",#每轮结束后进行评价
    save_strategy = "epoch",#每个epoch保存一次权重，默认是steps
    learning_rate=2e-5,#初始学习率
    per_device_train_batch_size=batch_size,#训练批次大小
    per_device_eval_batch_size=batch_size,#测试批次大小
    num_train_epochs=5,#训练轮数
    weight_decay=0.01,#指数衰减？
    load_best_model_at_end=True,#是否在训练结束时加载训练过程中找到的最佳模型。默认否
    metric_for_best_model=metric_name,#通过str方式传递评测方法。结合第一句表示stsb使用皮尔逊系数，cola使用matthews_correlation，其它都是acc
    #这个表示训练中评测效果更好就保存权重参数，否则继续训练但是不更新。相当于early-stop
    log_level='error',
    logging_strategy="no",
    report_to="none"
)

# 根据任务名称获取不同的评测方法
def compute_metrics(eval_pred):
    predictions, labels = eval_pred
    if task != "stsb":
        predictions = np.argmax(predictions, axis=1)
    else:
        predictions = predictions[:, 0]
    return metric.compute(predictions=predictions, references=labels)

# 构造训练器Trainer
validation_key = "validation_mismatched" if task == "mnli-mm" else "validation_matched" if task == "mnli" else "validation"
trainer = Trainer(
    model,
    args,
    train_dataset=encoded_dataset["train"],
    eval_dataset=encoded_dataset[validation_key],
    tokenizer=tokenizer,
    compute_metrics=compute_metrics
)

还有参数优化器，默认是Adamw。

4.3 训练模型

trainer.train()

TrainOutput(global_step=2675, training_loss=0.2717150308484229, metrics={'train_runtime': 100.5668, 'train_samples_per_second': 425.14, 'train_steps_per_second': 26.599, 'total_flos': 229537542078168.0, 'train_loss': 0.2717150308484229, 'epoch': 5.0})

4.4 模型评估

trainer.evaluate()

[66/66 00:00]

{'eval_loss': 0.8624260425567627,
 'eval_matthews_correlation': 0.519563286537562,
 'eval_runtime': 0.6501,
 'eval_samples_per_second': 1604.31,
 'eval_steps_per_second': 101.519,
 'epoch': 5.0}

5 超参数搜索

Trainer同样支持超参搜索，使用optuna or Ray Tune代码库。
反注释下面两行安装依赖：

! pip install optuna
! pip install ray[tune]

5.1 设置初始化模型

  超参搜索时，Trainer将会返回多个训练好的模型，所以需要传入一个定义好的模型从而让Trainer可以不断重新初始化该传入的模型：

def model_init():
    return AutoModelForSequenceClassification.from_pretrained(
    model_checkpoint, num_labels=num_labels)

#调用 Trainer
trainer = Trainer(
    model_init=model_init,
    args=args,
    train_dataset=encoded_dataset["train"],
    eval_dataset=encoded_dataset[validation_key],
    tokenizer=tokenizer,
    compute_metrics=compute_metrics
)

5.2 超参数搜索

  调用方法hyperparameter_search进行超参搜索。这个过程可能很久，故可以先用部分数据集进行超参搜索，再进行全量训练。 比如使用1/10的数据进行搜索：

# 使用1/10数据进行搜索
best_run = trainer.hyperparameter_search(n_trials=10, direction="maximize")

# hyperparameter_search会返回到效果最好的模型参数
best_run

BestRun(run_id='3', objective=0.5504031254980248, hyperparameters={'learning_rate': 4.301257551502102e-05, 'num_train_epochs': 5, 'seed': 20, 'per_device_train_batch_size': 8})

5.3 设置效果最好的参数并训练模型

将Trainner设置为搜索到的最好参数，进行训练：

for n, v in best_run.hyperparameters.items():
    setattr(trainer.args, n, v)
trainer.train()

TrainOutput(global_step=5345, training_loss=0.26719996967083726, metrics={'train_runtime': 178.4912, 'train_samples_per_second': 239.536, 'train_steps_per_second': 29.945, 'total_flos': 413547436355364.0, 'train_loss': 0.26719996967083726, 'epoch': 5.0})

trainer.evaluate()

    {'eval_loss': 0.9789257049560547,
     'eval_matthews_correlation': 0.5548273578107759,
     'eval_runtime': 0.6556,
     'eval_samples_per_second': 1590.796,
     'eval_steps_per_second': 100.664,
     'epoch': 5.0}

6 总结

  本次任务，主要介绍了用BERT模型解决文本分类任务的方法及步骤，步骤主要分为加载数据、数据预处理、微调预训练模型和超参数搜索。在加载数据阶段中，必须使用与分类任务相应的评测方法；在数据预处理阶段中，对tokenizer分词器的建模，并完成数据集中所有样本的预处理；在微调预训练模型阶段，通过对模型参数进行设置，并构建Trainner训练器，进行模型训练和评估；最后在超参数搜索阶段，使用hyperparameter_search方法，搜索效果最好的超参数，并进行模型训练和评估。
  其中在数据集下载时，需要使用外网方式建立代理。如果使用conda安装ray[tune]包时，请下载对应ray-tune依赖包。