前言 如果要实战，肯定第一个考虑的库就是transformers, 关于transformers的用法可以参考这个视频，是个外国的小哥, 很短很实用。
【双语字幕+资料下载】Hugging Face速成指南！一遍搞定NLP任务中最常用的功能板块<实战教程系列>
本文就是根据这个视频进行实验的。
首先，如果没有 transformers库的话，先执行

pip install transformers

另外，torch可以去官网下载(latest)，如果需要下载其他版本的话, 可以去https://download.pytorch.org/whl/torch_stable.html



Pipeline 这个仅用于一些常规任务，自己想简单处理就行，不求定制什么模型，以及不需要finetune的场景下使用。

from transformers import pipeline

classifier = pipeline("sentiment-analysis") res = classifier("We are very happy to show you the Transformers library.") print(res)

可以看到默认使用的模型是 distilbert-base-uncased-finetuned-sst-2-english， 输出结果:
[{‘label’: ‘POSITIVE’, ‘score’: 0.9997994303703308}]

文章图片


多个输入样例:

classifier = pipeline("sentiment-analysis") res = classifier(["We are very happy to show you the Transformers library.", "We hope you don't hate it."]) print(res)

输出:
[{‘label’: ‘POSITIVE’, ‘score’: 0.9997994303703308}, {‘label’: ‘NEGATIVE’, ‘score’: 0.5308617353439331}]



指定模型:

model_name = 'distilbert-base-uncased-finetuned-sst-2-english'classifier = pipeline("sentiment-analysis", model=model_name) res = classifier(["We are very happy to show you the Transformers library.", "We hope you don't hate it."]) print(res)

输出:
[{‘label’: ‘POSITIVE’, ‘score’: 0.9997994303703308}, {‘label’: ‘NEGATIVE’, ‘score’: 0.5308617353439331}]



定制分词器和模型

from transformers import pipeline from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_name = 'distilbert-base-uncased-finetuned-sst-2-english'tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForSequenceClassification.from_pretrained(model_name)classifier = pipeline("sentiment-analysis", model=model, tokenizer=tokenizer) res = classifier(["We are very happy to show you the Transformers library.", "We hope you don't hate it."]) print(res)

输出:
[{‘label’: ‘POSITIVE’, ‘score’: 0.9997994303703308}, {‘label’: ‘NEGATIVE’, ‘score’: 0.5308617353439331}]



tokenizer

from transformers import AutoTokenizer

这个是分词器，我们需要把文本转化为适合模型输入的向量。(bert家族的话是input_ids和attention_mask, 注意input_ids是不需要 onehot 处理的)

tokens = tokenizer.tokenize('We are very happy to show you the Transformers library.') token_ids = tokenizer.convert_tokens_to_ids(tokens)print(f'tokens: {tokens}') print(f'token_ids: {token_ids}')

tokens: [‘we’, ‘are’, ‘very’, ‘happy’, ‘to’, ‘show’, ‘you’, ‘the’, ‘transformers’, ‘library’, ‘.’]
token_ids: [2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 19081, 3075, 1012]
可以看到， tokenize方法就是分词, convert_tokens_to_ids就是分完的token转化为id。



来看下另一个例子，可以看到我们根据输出就可以发现，该方式的输出 即input是一个字典，该字典可以解包后作为model的输入。

input = tokenizer("We are very happy to show you the Transformers library.") print(input)

{‘input_ids’: [101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 19081, 3075, 1012, 102], ‘attention_mask’: [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}



一个batch多个样例, 其中padding是指填充的意思，例如第二句句子长度不足，则会补上0 (通常是0)进行填充。

train_x = ["We are very happy to show you the Transformers library.", "We hope you don't hate it."]batch = tokenizer(train_x, padding=True, truncation=True, max_length=512, return_tensors='pt') print(batch)

{‘input_ids’: tensor([[ 101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 19081,
3075, 1012, 102],
[ 101, 2057, 3246, 2017, 2123, 1005, 1056, 5223, 2009, 1012,
102, 0, 0]]), ‘attention_mask’: tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]])}


不指定返回是pytorch.tensor类型

train_x = ["We are very happy to show you the Transformers library.", "We hope you don't hate it."]batch = tokenizer(train_x, padding=True, truncation=True, max_length=512) print(batch)

{‘input_ids’: [[101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 19081, 3075, 1012, 102], [101, 2057, 3246, 2017, 2123, 1005, 1056, 5223, 2009, 1012, 102, 0, 0]], ‘attention_mask’: [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]}



另外一种就是encode, 可以看到它的输出是input_ids, 但是它会连在一起，是一个一维的输出， 所以使用encode的话，下面用法并不对， 得用for循环去遍历，具体可看情感分析bert家族 pytorch实现(ing

train_x = ["We are very happy to show you the Transformers library.", "We hope you don't hate it."]model_name = 'distilbert-base-uncased-finetuned-sst-2-english' tokenizer = AutoTokenizer.from_pretrained(model_name)batch = tokenizer.encode(train_x, padding=True, truncation=True, max_length=512) print(batch)

[101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 19081, 3075, 1012, 102, 2057, 3246, 2017, 2123, 1005, 1056, 5223, 2009, 1012, 102]



inference 利用model进行分类, 这里我们仅做推导，稍后介绍如何finetue。

model_name = 'distilbert-base-uncased-finetuned-sst-2-english'tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForSequenceClassification.from_pretrained(model_name)train_x = ["We are very happy to show you the Transformers library.", "We hope you don't hate it."]batch = tokenizer(train_x, padding=True, truncation=True, max_length=512, return_tensors='pt')with torch.no_grad(): outputs = model(**batch) predictions = F.softmax(outputs.logits, dim=-1) labels = torch.argmax(predictions, dim=-1) labels_en = [model.config.id2label[label_id] for label_id in labels.tolist()]print('outputs:\n', outputs) print('predictions:\n', predictions) print('labels:', labels) print('labels_en:', labels_en)

outputs:
SequenceClassifierOutput(loss=None, logits=tensor([[-4.1329, 4.3811],
[ 0.0818, -0.0418]]), hidden_states=None, attentions=None)
predictions:
tensor([[2.0060e-04, 9.9980e-01],
[5.3086e-01, 4.6914e-01]])
labels: tensor([1, 0])
labels_en: [‘POSITIVE’, ‘NEGATIVE’]



【人工智能|hugginface-introduction 案例介绍】这里我们注意到上面的loss没有损失，这里我们再添加一个ground truth

# 有损失 model_name = 'distilbert-base-uncased-finetuned-sst-2-english'tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForSequenceClassification.from_pretrained(model_name)train_x = ["We are very happy to show you the Transformers library.", "We hope you don't hate it."] train_y = [1, 0]batch = tokenizer(train_x, padding=True, truncation=True, max_length=512, return_tensors='pt')with torch.no_grad(): outputs = model(**batch, labels=torch.tensor(train_y)) predictions = F.softmax(outputs.logits, dim=-1) labels = torch.argmax(predictions, dim=-1) labels_en = [model.config.id2label[label_id] for label_id in labels.tolist()]print('outputs:\n', outputs) print('predictions:\n', predictions) print('labels:', labels) print('labels_en:', labels_en)

outputs:
SequenceClassifierOutput(loss=tensor(0.3167), logits=tensor([[-4.1329, 4.3811],
[ 0.0818, -0.0418]]), hidden_states=None, attentions=None)
predictions:
tensor([[2.0060e-04, 9.9980e-01],
[5.3086e-01, 4.6914e-01]])
labels: tensor([1, 0])
labels_en: [‘POSITIVE’, ‘NEGATIVE’]



这里其实tokenizer不是返回的字典里的元素不是torch.tensor类型也可以

# 有损失 model_name = 'distilbert-base-uncased-finetuned-sst-2-english'tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForSequenceClassification.from_pretrained(model_name)train_x = ["We are very happy to show you the Transformers library.", "We hope you don't hate it."]batch_dict = tokenizer(train_x, padding=True, truncation=True, max_length=512) print('batch_dict:', batch_dict) batch = torch.tensor(batch_dict['input_ids']) print('batch:', batch_dict)with torch.no_grad(): outputs = model(batch) predictions = F.softmax(outputs.logits, dim=-1) labels = torch.argmax(predictions, dim=-1) labels_en = [model.config.id2label[label_id] for label_id in labels.tolist()]print('outputs:\n', outputs) print('predictions:\n', predictions) print('labels:', labels) print('labels_en:', labels_en)

batch_dict: {‘input_ids’: [[101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 19081, 3075, 1012, 102], [101, 2057, 3246, 2017, 2123, 1005, 1056, 5223, 2009, 1012, 102, 0, 0]], ‘attention_mask’: [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]}
batch: {‘input_ids’: [[101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 19081, 3075, 1012, 102], [101, 2057, 3246, 2017, 2123, 1005, 1056, 5223, 2009, 1012, 102, 0, 0]], ‘attention_mask’: [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]}
outputs:
SequenceClassifierOutput(loss=None, logits=tensor([[-4.1329, 4.3811],
[ 1.5112, -1.3358]]), hidden_states=None, attentions=None)
predictions:
tensor([[2.0060e-04, 9.9980e-01],
[9.4517e-01, 5.4834e-02]])
labels: tensor([1, 0])
labels_en: [‘POSITIVE’, ‘NEGATIVE’]



加上label, 我们关注一下loss, 可以看到没有传入attention_mask其实loss的结果是不同的，尽管输出的标签一样

# 有损失 model_name = 'distilbert-base-uncased-finetuned-sst-2-english'tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForSequenceClassification.from_pretrained(model_name)train_x = ["We are very happy to show you the Transformers library.", "We hope you don't hate it."] train_y = [1, 0]batch_dict = tokenizer(train_x, padding=True, truncation=True, max_length=512) print('batch_dict:', batch_dict) batch = torch.tensor(batch_dict['input_ids']) print('batch:', batch_dict)with torch.no_grad(): outputs = model(batch, labels=torch.tensor(train_y)) predictions = F.softmax(outputs.logits, dim=-1) labels = torch.argmax(predictions, dim=-1) labels_en = [model.config.id2label[label_id] for label_id in labels.tolist()]print('outputs:\n', outputs) print('predictions:\n', predictions) print('labels:', labels) print('labels_en:', labels_en)

batch_dict: {‘input_ids’: [[101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 19081, 3075, 1012, 102], [101, 2057, 3246, 2017, 2123, 1005, 1056, 5223, 2009, 1012, 102, 0, 0]], ‘attention_mask’: [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]}
batch: {‘input_ids’: [[101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 19081, 3075, 1012, 102], [101, 2057, 3246, 2017, 2123, 1005, 1056, 5223, 2009, 1012, 102, 0, 0]], ‘attention_mask’: [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]}
outputs:
SequenceClassifierOutput(loss=tensor(0.0283), logits=tensor([[-4.1329, 4.3811],
[ 1.5112, -1.3358]]), hidden_states=None, attentions=None)
predictions:
tensor([[2.0060e-04, 9.9980e-01],
[9.4517e-01, 5.4834e-02]])
labels: tensor([1, 0])
labels_en: [‘POSITIVE’, ‘NEGATIVE’]



save & load 保存模型

import os from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_name = 'distilbert-base-uncased-finetuned-sst-2-english'tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForSequenceClassification.from_pretrained(model_name)

保存权重, 到指定文件夹

save_dir = 'save_dir'if os.path.exists(save_dir): os.mkdir(save_dir)tokenizer.save_pretrained(save_dir) model.save_pretrained(save_dir)

文章图片


下次进行加载

load_tokenizer = AutoTokenizer.from_pretrained(save_dir) load_model = AutoModelForSequenceClassification.from_pretrained(save_dir)

fine-tune

from pathlib import Path from sklearn.model_selection import train_test_split import torch from torch.utils.data import Dataset from transformers import DistilBertTokenizer, DistilBertForSequenceClassification from transformers import Trainer, TrainingArguments

def read_imdb_split(split_dir): split_dir = Path(split_dir) texts = [] labels = [] for label_dir in ['pos', 'neg']: for text_file in (split_dir/label_dir).iterdir(): texts.append(text_file.read_text()) labels.append(0 if label_dir == 'neg' else 1) return texts, labels

下载数据集

# http://ai.stanford.edu/~amaas/data/sentiment !wget https://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz

解压数据集

!tar -zxvf /content/aclImdb_v1.tar.gz

文章图片


读取数据集

train_texts, train_labels = read_imdb_split('/content/aclImdb/train') test_texts, test_labels = read_imdb_split('/content/aclImdb/test') print('train_texts.len:', len(train_texts)) print('test_texts.len: ', len(test_texts))

train_texts.len: 25000
test_texts.len: 25000



分一部分作于验证

train_texts, val_texts, train_labels, val_labels = train_test_split(train_texts, train_labels, test_size=0.2) print('train_texts.len:', len(train_texts)) print('val_texts.len: ', len(val_texts))

train_texts.len: 12800
val_texts.len: 3200

class IMDbDataset(Dataset): def __init__(self, encodings, labels): self.encodings = encodings self.labels = labelsdef __getitem__(self, idx): item = {key:torch.tensor(val[idx]) for key, val in self.encodings.items()} item['labels'] = torch.tensor(self.labels[idx]) return itemdef __len__(self): return len(self.labels)

model_name = 'distilbert-base-uncased-finetuned-sst-2-english'tokenizer = DistilBertTokenizer.from_pretrained(model_name)train_encodings = tokenizer(train_texts, truncation=True, padding=True) val_encodings = tokenizer(val_texts, truncation=True, padding=True) test_encodings = tokenizer(test_texts, truncation=True, padding=True)

train_dataset = IMDbDataset(train_encodings, train_labels) val_dataset = IMDbDataset(val_encodings, val_labels) test_dataset = IMDbDataset(test_encodings, test_labels)

trainer写法 设置训练的超参数

training_args = TrainingArguments( output_dir = './results', num_train_epochs=2, per_device_train_batch_size=16, per_device_eval_batch_size=64, warmup_steps=500, learning_rate=5e-5, weight_decay=0.01, logging_dir='./logs', logging_steps=10 )

加载模型

model = DistilBertForSequenceClassification.from_pretrained(model_name)

加载训练器

trainer = Trainer( model=model, args=training_args, train_dataset=train_dataset, eval_dataset=val_dataset )

开始训练

trainer.train()

文章图片



native pytorch写法

from torch.utils.data import DataLoader from transformers import AdamW

device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') print(device)

加载模型

model = DistilBertForSequenceClassification.from_pretrained(model_name) model.to(device) model.train()

定义数据Loader类

train_loader = DataLoader(train_dataset, batch_size=16, shuffle=True)

实例化优化器和设置epoch数量

optim = AdamW(model.parameters(), lr=5e-5) num_train_epochs = 2

for epoch in range(num_train_epochs): for batch in train_loader: optim.zero_grad() input_ids = batch['input_ids'].to(device) attention_mask = batch['attention_mask'].to(device) labels = batch['labels'].to(device)outputs = model(input_ids, attention_mask=attention_mask, labels=labels)loss = outputs[0] print('epoch:{}, batch:{}, loss:{}'.format(epoch, batch, loss)) loss.backward() optim.step()

文章图片

推荐阅读

• 

  智慧保险的快速发展,未来的保险公司将是什么模样呢？
• 

  获取oss的pdf文件，返回文件输出流
• 

  2023重庆涪陵3月赏花时间+地点 涪陵的花
• 

  wnda是什么轮胎 wnda是什么轮胎里面有一层胶
• 

  时光与岁月优美短句 关于时光岁月的唯美句子
• 

  mysql怎么设置字段默认值为0 mysql设置字段为小时
• 

  魔兽世界桃乐豆详细介绍
• 

  食色,性也究竟是什么意思 食色性也是什么意思
• 

  什么材质的行李箱好一点 什么材质的行李箱好
• 

  华为网盘会被和谐吗,华为网盘有什么用？
• 

  成都徒步旅行～懒鱼第四天作业
• 

  关系胜于一切，关系就是温暖
• 

  g三星 Galaxy S22 / S22 +图赏：老设计焕发新感觉
• 

  北京医保提到4000北京医保提到4000的原因（北京医保提到4000北京医保
• 

  文字代码生成器，谁有日志彩字生成代码器
• 

  在网店转让平台上如何让自己的店铺快速出售？有哪些小技巧？
• 

  十七岁，我曾有一个朋友
• 

  潇潇雨歇是什么意思 潇潇雨歇解释
• 

  四轴数控车床怎么编程，请大师教教我这个用四轴怎么编程
• 

  怎么判断水果是否有毒

• 人脸识别|【人脸识别系列】| 实现自动化妆
• 人工智能|干货！人体姿态估计与运动预测
• 推荐系统论文进阶|CTR预估 论文精读(十一)--Deep Interest Evolution Network(DIEN)
• Python专栏|数据分析的常规流程
• 历史上的今天|【历史上的今天】2 月 16 日（世界上第一个 BBS 诞生；中国计算机教育开端；IBM 机器人赢得智能竞赛）
• 网络|一文彻底搞懂前端监控
• 游戏|2022年如何学习前端前沿技术，破卷而出（）
• 分布式|《Python3网络爬虫开发实战（第二版）》内容介绍