「Huggingface🤗NLP笔记系列-第7集」 最近跟着Huggingface上的NLP tutorial走了一遍，惊叹居然有如此好的讲解Transformers系列的NLP教程，于是决定记录一下学习的过程，分享我的笔记，可以算是官方教程的精简+注解版。但最推荐的，还是直接跟着官方教程来一遍，真是一种享受。

• 官方教程网址：https://huggingface.co/course/chapter1
• 本期内容对应网址：https://huggingface.co/course/chapter3/3?fw=pt
• 本系列笔记的GitHub： https://github.com/beyondguo/Learn_PyTorch/tree/master/HuggingfaceNLP

---

使用Trainer API来微调模型

1. 数据集准备和预处理：

这部分就是回顾上一集的内容：

• 通过dataset包加载数据集
• 加载预训练模型和tokenizer
• 定义Dataset.map要使用的预处理函数
• 定义DataCollator来用于构造训练batch

import numpy as np
from transformers import AutoTokenizer, DataCollatorWithPadding
import datasets
checkpoint = 'bert-base-cased'
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
raw_datasets = datasets.load_dataset('glue', 'mrpc')

def tokenize_function(sample):
    return tokenizer(sample['sentence1'], sample['sentence2'], truncation=True)
tokenized_datasets = raw_datasets.map(tokenize_function, batched=True)

data_collator = DataCollatorWithPadding(tokenizer=tokenizer)

2. 加载我们要fine-tune的模型：

from transformers import AutoModelForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=2)

>>> (warnings)
Some weights of the model checkpoint at bert-base-cased were not used when initializing BertForSequenceClassification: ['cls.predictions.bias', 'cls.predictions.transform.dense.weight', 'cls.predictions.transform.dense.bias', 'cls.predictions.decoder.weight', 'cls.seq_relationship.weight', 'cls.seq_relationship.bias', 'cls.predictions.transform.LayerNorm.weight', 'cls.predictions.transform.LayerNorm.bias']
- This IS expected if you are initializing BertForSequenceClassification 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 BertForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
Some weights of BertForSequenceClassification were not initialized from the model checkpoint at bert-base-cased and are newly initialized: ['classifier.weight', 'classifier.bias']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.

不得不说，这个Huggingface很贴心，这里的warning写的很清楚。这里我们使用的是带ForSequenceClassification这个Head的模型，但是我们的bert-baed-cased虽然它本身也有自身的Head，但跟我们这里的二分类任务不匹配，所以可以看到，它的Head被移除了，使用了一个随机初始化的ForSequenceClassificationHead。

所以这里提示还说："You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference."

3. 使用Trainer来训练

Trainer是Huggingface transformers库的一个高级API，可以帮助我们快速搭建训练框架：

from transformers import Trainer, TrainingArguments

training_args = TrainingArguments(output_dir='test_trainer') # 指定输出文件夹，没有会自动创建

trainer = Trainer(
    model,
    training_args,
    train_dataset=tokenized_datasets["train"],
    eval_dataset=tokenized_datasets["validation"],
    data_collator=data_collator,  # 在定义了tokenizer之后，其实这里的data_collator就不用再写了，会自动根据tokenizer创建
    tokenizer=tokenizer,
)

我们看看TrainingArguments和Trainer的参数都有些啥：

• https://huggingface.co/transformers/master/main_classes/trainer.html
• https://huggingface.co/transformers/master/main_classes/trainer.html#trainingarguments

TrainingArguments(
    output_dir: Union[str, NoneType] = None,
    overwrite_output_dir: bool = False,
    do_train: bool = False,
    do_eval: bool = None,
    do_predict: bool = False,
    evaluation_strategy: transformers.trainer_utils.EvaluationStrategy = 'no',
    prediction_loss_only: bool = False,
    per_device_train_batch_size: int = 8,  # 默认的batch_size=8
    per_device_eval_batch_size: int = 8,
    per_gpu_train_batch_size: Union[int, NoneType] = None,
    per_gpu_eval_batch_size: Union[int, NoneType] = None,
    gradient_accumulation_steps: int = 1,
    eval_accumulation_steps: Union[int, NoneType] = None,
    learning_rate: float = 5e-05,
    weight_decay: float = 0.0,
    adam_beta1: float = 0.9,
    adam_beta2: float = 0.999,
    adam_epsilon: float = 1e-08,
    max_grad_norm: float = 1.0,
    num_train_epochs: float = 3.0,   # 默认跑3轮
    ...

Trainer(
    model: Union[transformers.modeling_utils.PreTrainedModel, torch.nn.modules.module.Module] = None,
    args: transformers.training_args.TrainingArguments = None,
    data_collator: Union[DataCollator, NoneType] = None,
    train_dataset: Union[torch.utils.data.dataset.Dataset, NoneType] = None,
    eval_dataset: Union[torch.utils.data.dataset.Dataset, NoneType] = None,
    tokenizer: Union[ForwardRef('PreTrainedTokenizerBase'), NoneType] = None,
    model_init: Callable[[], transformers.modeling_utils.PreTrainedModel] = None,
    compute_metrics: Union[Callable[[transformers.trainer_utils.EvalPrediction], Dict], NoneType] = None,
    callbacks: Union[List[transformers.trainer_callback.TrainerCallback], NoneType] = None,
    optimizers: Tuple[torch.optim.optimizer.Optimizer, torch.optim.lr_scheduler.LambdaLR] = (None, None),  # 默认会使用AdamW
)
Docstring:     
Trainer is a simple but feature-complete training and eval loop for PyTorch, optimized for 🤗 Transformers.

可见，这个Trainer把所有训练中需要考虑的参数、设计都包括在内了，我们可以在这里指定训练验证集、data_collator、metrics、optimizer，并通过TrainingArguments来提供各种超参数。

默认情况下，Trainer和TrainingArguments会使用：

• batch size=8
• epochs = 3
• AdamW优化器

定义好之后，直接使用.train()来启动训练：

trainer.train()

输出：

TrainOutput(global_step=1377, training_loss=0.35569445984728887, metrics={'train_runtime': 383.0158, 'train_samples_per_second': 3.595, 'total_flos': 530185443455520, 'epoch': 3.0})

然后我们用Trainer来预测：

trainer.predict()函数处理的结果是一个named_tuple（一种可以直接通过key来取值的tuple），类似一个字典，包含三个属性：predictions, label_ids, metrics

注意，这里的三个属性：

• predictions实际上就是logits
• label_ids不是预测出来的id，而是数据集中自带的ground truth的标签，因此如果输入的数据集中没给标签，这里也不会输出
• metrics，也是只有输入的数据集中提供了label_ids才会输出metrics，包括loss之类的指标

其中metrics中还可以包含我们自定义的字段，我们需要在定义Trainer的时候给定compute_metrics参数。

文档参考： https://huggingface.co/transformers/master/main_classes/trainer.html#transformers.Trainer.predict

predictions = trainer.predict(tokenized_datasets['validation'])
print(predictions.predictions.shape)  # logits
# array([[-2.7887206,  3.1986978],
#       [ 2.5258656, -1.832253 ], ...], dtype=float32)
print(predictions.label_ids.shape) # array([1, 0, 0, 1, 0, 1, 0, 1, 1, 1, ...], dtype=int64)
print(predictions.metrics)

输出：

[51/51 00:03]
(408, 2)
(408,)
{'eval_loss': 0.7387174963951111, 'eval_runtime': 3.2872, 'eval_samples_per_second': 124.117}

然后就可以用preds和labels来计算一些相关的metrics了。

Huggingface datasets里面可以直接导入跟数据集相关的metrics：

from datasets import load_metric

preds = np.argmax(predictions.predictions, axis=-1)

metric = load_metric('glue', 'mrpc')
metric.compute(predictions=preds, references=predictions.label_ids)

>>>
{'accuracy': 0.8455882352941176, 'f1': 0.8911917098445595}

看看这里的metric（glue type）的文档：

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

4.构建Trainer中的compute_metrics函数

前面我们注意到Trainer的参数中，可以提供一个compute_metrics函数，用于输出我们希望有的一些指标。

这个compute_metrics有一些输入输出的要求：

• 输入：是一个EvalPrediction对象，是一个named tuple，需要有至少predictions和label_ids两个字段；经过查看源码，这里的predictions，就是logits
• 输出：一个字典，包含各个metrics和对应的数值。

源码地址： https://huggingface.co/transformers/master/_modules/transformers/trainer.html#Trainer

from datasets import load_metric
def compute_metrics(eval_preds):
    metric = load_metric("glue", "mrpc")
    logits, labels = eval_preds.predictions, eval_preds.label_ids
    # 上一行可以直接简写成：
    # logits, labels = eval_preds  因为它相当于一个tuple
    predictions = np.argmax(logits, axis=-1)
    return metric.compute(predictions=predictions, references=labels)

总结一下这个过程：

• 首先我们定义了一个compute_metrics函数，交给Trainer；
• Trainer训练模型，模型会对样本计算，产生 predictions (logits)；
• Trainer再把 predictions 和数据集中给定的 label_ids 打包成一个对象，发送给compute_metrics函数；
• compute_metrics函数计算好相应的 metrics 然后返回。

看看带上了 compute_metrics 之后的训练：

training_args = TrainingArguments(output_dir='test_trainer', evaluation_strategy='epoch')
model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=2)  # new model
trainer = Trainer(
    model,
    training_args,
    train_dataset=tokenized_datasets["train"],
    eval_dataset=tokenized_datasets["validation"],
    data_collator=data_collator,  # 在定义了tokenizer之后，其实这里的data_collator就不用再写了，会自动根据tokenizer创建
    tokenizer=tokenizer,
    compute_metrics=compute_metrics
)

trainer.train()

输出：

TrainOutput(global_step=1377, training_loss=0.32063739751678666, metrics={'train_runtime': 414.1719, 'train_samples_per_second': 3.325, 'total_flos': 530351810395680, 'epoch': 3.0})

可见，带上了compute_metircs函数之后，在Trainer训练过程中，会把增加的metric也打印出来，方便我们时刻了解训练的进展。