Coggle 30 Days of ML（23年7月）打卡

code_mryxj

已于 2023-07-13 23:57:52 修改

阅读量281

点赞数

文章标签： nlp bert

于 2023-07-13 19:16:03 首次发布

本文链接：https://blog.csdn.net/yexiaohhjk/article/details/131709289

版权

前言

最近开始关注LLM相关知识，但之前的NLP范式的技能不能丢。
这个练习还是比较适合我，感谢举办方选题，快速全部打卡一波。

打卡记录

任务一: 报名比赛，下载比赛数据集并完成读取

比赛链接：https://challenge.xfyun.cn/topic/info?type=text-detector&ch=vWxQGFU
报名，下载数据到本地后，利用pandas库读取csv文件，

import pandas as pd

def load_data(data_dir='../data/'):
    '''ChatGPT生成文本检测器公开数据.zip
    '''
    train = pd.read_csv(data_dir + 'train.csv', 
            sep=',', names=['name', 'label', 'content'])
    
    test = pd.read_csv(data_dir + 'test.csv', 
            sep=',', names=['name', 'content'])
    
    return train, test

train, test = load_data()

在notebook输出数据，看出文本已经通过一层映射脱敏处理处理成连续的token id，意味着后面无法直接用现成pretrain 好的model或者embedding进行微调，需要重新预训练一下。

train

	name	label	content
0	name	label	content
1	1	0	[4509 3181 1253 2278 290 3562 2051 599 3125 ...
2	2	1	[ 263 1325 2563 4160 2196 169 3125 2563 2619 ...
3	3	0	[3635 177 3125 1251 3839 5212 2109 1171 1194 ...
4	4	1	[3037 266 246 3547 1253 2278 3125 649 697 ...
...	...	...	...
13996	13996	0	[5212 1759 1953 139 1953 3180 3187 5212 3414 ...
13997	13997	1	[ 699 778 2777 1333 224 3543 998 139 4411 ...
13998	13998	1	[ 506 211 139 3333 3293 286 4358 272 5212 ...
13999	13999	1	[1583 169 123 2969 998 5212 1759 266 1435 ...
14000	14000	0	[1759 266 4399 205 5212 1759 266 4399 205 ...

14001 rows × 3 columns

test

	name	content
0	name	content
1	14001	[3125 2196 286 123 1539 1759 266 3549 649 ...
2	14002	[1109 2113 3122 213 3125 1294 5212 2338 2233 ...
3	14003	[ 236 3125 139 3037 5212 4294 1600 4550 3169 ...
4	14004	[ 13 13 13 0 0 0 245 1472 3125 ...
...	...	...
9996	23996	[ 430 4829 302 3447 5212 2081 2176 1146 2321 ...
9997	23997	[ 0 109 3324 532 1294 199 1370 2176 1105 ...
9998	23998	[3406 1093 5212 3036 3635 3125 3406 1296 5212 ...
9999	23999	[ 0 0 3799 248 184 5195 4399 205 434 ...
10000	24000	[2471 4553 262 1759 2579 4555 1560 4549 3125 ...

10001 rows × 2 columns

任务二：对数据集字符进行可视化，统计标签和字符分布

对数据集中train,test集合分别统计基本特征，看一下数据特征的分布是否一致。

这里数据是文本，主要特征就是文本的长度分布。

import pandas as pd

def plot_data_len_distribution(data_lens,
                               bins=None,
                               columns=['区间的数量']):
    if bins:
        cats = pd.cut(data_lens, bins)
        value_counts = pd.value_counts(cats, sort=False)
    else:
        value_counts = pd.value_counts(data_lens, sort=False)
    df = pd.DataFrame(value_counts, columns=columns)
    df.plot(kind='bar')     
    df.plot.pie(subplots=True, autopct="%.2f%%")
    

train = train[1:]
test = test[1:]
train['label'] = train['label'].apply(lambda x: int(x))
test['label'] = train['label'].apply(lambda x: int(x))
train['content_len'] = train['content'].apply(lambda x:x[1:-1].split()).apply(lambda x: len(x))
test['content_len'] = test['content'].apply(lambda x:x[1:-1].split()).apply(lambda x: len(x))

plot_data_len_distribution(train['label'].tolist(),
                           columns=['标签的数量'])
    
plot_data_len_distribution(train['content_len'].tolist(),
                           bins = [180, 200, 230, 250, 260, 270, 275, 280, 290, 300, 320, 350, 360, 500, 1000], 
                           columns=['字符区间的数量'])

plot_data_len_distribution(test['content_len'].tolist(),
                           bins = [180, 200, 230, 250, 260, 270, 275, 280, 290, 300, 320, 350, 360, 500, 1000],
                           columns=['字符区间的数量'])

统计训练数据中label标签分布：0，1标签数量比大概 6 ：1，如下:
在这里插入图片描述

训练数据中文本长度的分布如下：

在这里插入图片描述

测试数据中文本长度的分布和训练集一致，文本长度都是200。

主要可以看出两点：

训练集中正负样本1：6
训练集和测试集文本长度全是200

任务三：使用TFIDF提取文本特征

import numpy as np
import pandas as pd

from sklearn.feature_extraction.text import TfidfVectorizer


train['content'] = train['content'].apply(lambda x: x[1:-1])
test['content'] = test['content'].apply(lambda x: x[1:-1])

train['label'] = train['label'].apply(lambda x: int(x))

# https://scikit-learn.org/stable/modules/generated/sklearn.feature_extraction.text.TfidfVectorizer.html
tfidf = TfidfVectorizer(token_pattern=r'\S+|\n',     # 数据中文本特征token被映射成token id，对token id采取整体。
                        analyzer='word',             
                        ngram_range=(1, 3),
                        max_features=2000)

train_tfidf = tfidf.fit_transform(train['content'])
test_tfidf = tfidf.fit_transform(test['content'])

任务四：使用TFIDF特征和线性模型完成训练和预测

为了验证不同模型方案效果，我在训练集进一步划分10%数据作为测试集

from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression

tfidf = TfidfVectorizer(token_pattern=r'\S+|\n',
                        analyzer='word',             
                        ngram_range=(1, 3), # ngram len 范围，调试时选取最佳
                        max_features=500)   # 调试时，选取最佳

# 在train集合上进一步划分train、test集合
train_more, test_more = train_test_split(train, 
                                         test_size=0.1, 
                                         random_state=43) 
train_more_tfidf = tfidf.fit_transform(train_more['content'])
test_more_tfidf = tfidf.fit_transform(test_more['content'])


# 使用Sklearn中的线性模型（如逻辑回归）进行训练，并使用训练好的模型对测试集进行预测。
lr_model = LogisticRegression()
lr_model.fit(
    train_more_tfidf,
    train_more['label']
)
test_more['predict'] = lr_model.predict(test_more_tfidf).astype(int)


# 评估模型的性能，如准确率、精确率、召回率等指标。
from sklearn.metrics import precision_score
from sklearn.metrics import accuracy_score
from sklearn.metrics import recall_score
from sklearn.metrics import f1_score
from sklearn.metrics import roc_curve
from sklearn.metrics import auc

pre = precision_score(test_more['label'], test_more['predict'])
acc = accuracy_score(test_more['label'], test_more['predict'])
rec = recall_score(test_more['label'], test_more['predict'])
f1 = f1_score(test_more['label'], test_more['predict'])

fpr, tpr, _  = roc_curve(test_more['label'], test_more['predict'])
auc = auc(fpr, tpr)

print(" pre: %.2f\n acc: %.2f\n rec: %.2f\n f1: %.2f\n auc: %.2f" % (pre, acc, rec, f1, auc))

 pre: 0.95
 acc: 0.96
 rec: 0.79
 f1: 0.86
 auc: 0.89

TFIDF + LR 提交结果

from sklearn.linear_model import LogisticRegression
from sklearn.feature_extraction.text import TfidfVectorizer


tfidf = TfidfVectorizer(token_pattern=r'\S+|\n',
                        analyzer='word',             
                        ngram_range=(1, 3),
                        max_features=500)

train_tfidf = tfidf.fit_transform(train['content'])
test_tfidf = tfidf.fit_transform(test['content'])

lr_model = LogisticRegression()
lr_model.fit(
    train_tfidf,
    train['label']
)
submit = pd.read_csv('./data/sample_submit.csv')
submit = submit.sort_values(by='name')
submit['label'] = lr_model.predict(test_tfidf).astype(int)
submit.to_csv('lr.csv', index=None)                               #提交大概 0.9223

任务五：使用TFIDF特征和XGBoost完成训练和预测

相比任务四，将lr替换成xgboost即可，效果有一定的提升，同时xgboost需要调试的参数更多一些。

from xgboost import XGBClassifier
from sklearn.model_selection import train_test_split
from sklearn.feature_extraction.text import TfidfVectorizer


tfidf = TfidfVectorizer(token_pattern=r'\S+|\n',
                        analyzer='word',             
                        ngram_range=(1, 3), # ngram len 范围，调试时选取最佳
                        max_features=500)   # 调试时，选取最佳

train['label'] = train['label'].apply(lambda x: int(x))
# 在train集合上进一步划分train、test集合
train_more, test_more = train_test_split(train, 
                                         test_size=0.1, 
                                         random_state=43) 
train_more_tfidf = tfidf.fit_transform(train_more['content'])
test_more_tfidf = tfidf.fit_transform(test_more['content'])


# 使用XGBoost进行训练，并使用训练好的模型对测试集进行预测。
xgb_model = XGBClassifier(max_depth=8, 
                          learning_rate=0.01, 
                          n_estimators=2000,
                          objective='binary:logistic',
                          booster='gbtree',
                          n_jobs=-1, 
                          nthread=None, 
                          gamma=0.3, 
                          min_child_weight=1, 
                          max_delta_step=0,
                          subsample=0.8, 
                          colsample_bytree=0.8,
                          colsample_bylevel=0.8, 
                          reg_alpha=0.5,
                          reg_lambda=0.5, 
                          scale_pos_weight=6,    # negative:postive = 6 :1 正样本权重
                          base_score=0.15,       # negative:postive = 6 :1 200/1400 = 0.14
                          random_state=233, 
                          seed=None)

xgb_model.fit(
    train_more_tfidf,
    train_more['label']
)
test_more['predict'] = xgb_model.predict(test_more_tfidf).astype(int)


# 评估模型的性能，如准确率、精确率、召回率等指标。
from sklearn.metrics import precision_score
from sklearn.metrics import accuracy_score
from sklearn.metrics import recall_score
from sklearn.metrics import f1_score
from sklearn.metrics import roc_curve
from sklearn.metrics import auc

pre = precision_score(test_more['label'], test_more['predict'])
acc = accuracy_score(test_more['label'], test_more['predict'])
rec = recall_score(test_more['label'], test_more['predict'])
f1 = f1_score(test_more['label'], test_more['predict'])
fpr, tpr, _  = roc_curve(test_more['label'], test_more['predict'])
auc = auc(fpr, tpr)

print(" pre: %.2f\n acc: %.2f\n rec: %.2f\n f1: %.2f\n auc: %.2f" % (pre, acc, rec, f1, auc))

 pre: 0.95
 acc: 0.94
 rec: 0.65
 f1: 0.77
 auc: 0.82

TFIDF + XGBoost 提交结果

from xgboost import XGBClassifier

tfidf = TfidfVectorizer(token_pattern=r'\S+|\n',
                        analyzer='word',             
                        ngram_range=(1, 3),
                        max_features=500)

train_tfidf = tfidf.fit_transform(train['content'])
test_tfidf = tfidf.fit_transform(test['content'])


# 使用XGBoost进行训练，并使用训练好的模型对测试集进行预测。
xgb_model = XGBClassifier(max_depth=8, 
                          learning_rate=0.01, 
                          n_estimators=2000,
                          objective='binary:logistic',
                          booster='gbtree',
                          n_jobs=-1, 
                          nthread=None, 
                          gamma=0.3, 
                          min_child_weight=1, 
                          max_delta_step=0,
                          subsample=0.8, 
                          colsample_bytree=0.8,
                          colsample_bylevel=0.8, 
                          reg_alpha=0.5,
                          reg_lambda=0.5, 
                          scale_pos_weight=6,    # negative:postive = 6 :1 正样本权重
                          base_score=0.15,       # negative:postive = 6 :1 200/1400 = 0.14
                          random_state=233, 
                          seed=None)

xgb_model.fit(
    train_tfidf,
    train['label']
)

submit = pd.read_csv('./data/sample_submit.csv')
submit = submit.sort_values(by='name')
submit['label'] = xgb_model.predict(test_tfidf).astype(int)
submit.to_csv('xgb.csv', index=None)                        # 0.9737

任务六：学会训练FastText、Word2Vec词向量

这一步利用gensim库进行word2vec训练词向量的步骤。

# 参考 https://radimrehurek.com/gensim/models/word2vec.html
# 以这个比赛数据为例, 准备语料
train, test = load_data()
train = train[1:]  # 第一行是属性名称
test = test[1:]

match_texts = []
for data in train["content"].tolist():
    for line in data[1:-1].split("\n"):
        match_texts.append(line.split(" "))
        
for data in test["content"].tolist():
    for line in data[1:-1].split("\n"):
        match_texts.append(line.split(" "))
        
print("该比赛数据集的语料一共有 %d 个句子" % len(match_texts))


#使用gensim库中的FastText类，设置相应的参数（如词向量维度、窗口大小、训练迭代次数等）来构建词向量模型。
from gensim.models import FastText
fasttext_model = FastText(vector_size=100,  
                          window=5,     
                          epochs=100,
                          min_count=2) 

# 使用gensim库中的Word2Vec类，设置相应的参数（如词向量维度、窗口大小、训练迭代次数等）来构建词向量模型。
from gensim.models.word2vec import Word2Vec
w2v_model = Word2Vec(vector_size=100,     # word embedding 维度
                     window=5,            # 窗口大小
                     epochs=100,          # 训练迭代次数
                     min_count=2,         # word 最低出现次数
                     sg=1,                # 1 skip-gram 0 cbow
                     workers=8)              

# 使用Word2Vec类的build_vocab()方法，构建词汇表
w2v_model.build_vocab(corpus_iterable=match_texts)

# 使用Word2Vec类的train()方法，训练词向量模型
w2v_model.train(corpus_iterable=match_texts,
                total_examples=len(match_texts), 
                epochs=100) 
w2v_model.save("../tmp/match_w2v_model.bin")
w2v_embed = w2v_model.wv
w2v_embed.save("../tmp/match_w2v_embed.bin")

def add_special_token(embed_path="", save_path=""):
    print("Loading [%s] embeddding." % (embed_path))
    wv_from_text = gensim.models.KeyedVectors.load(embed_path)
    vocab = wv_from_text.key_to_index
    print("Vocabulary Size: %s" % len(vocab.keys()))
    word_vocab = dict()
    word_vocab["PAD"] = 0
    word_vocab["UNK"] = 1
    for key in vocab.keys():
        word_vocab[key] = len(word_vocab.keys())
    pd.to_pickle(word_vocab, save_path+"vocab.pkl")
    word_embed = wv_from_text.vectors
    embed_dimen = word_embed.shape[-1]
    unk_embed = np.random.randn(1, embed_dimen)
    pad_embed = np.zeros(shape=(1, embed_dimen), dtype=np.float)
    extral_embed = np.concatenate((pad_embed, unk_embed), axis=0)
    word_embed = np.concatenate((extral_embed, word_embed), axis=0)
    np.save(save_path+"match_embed.npy", word_embed)
    print("Save add [Pad]、[UNK] Embedding shape: {}".format(word_embed.shape))
    word2idx = {word: idx for idx, word in enumerate(word_vocab)}
    idx2word = {idx: word for idx, word in enumerate(word_vocab)}
    pd.to_pickle(word2idx, save_path+"word2idx.pkl")
    pd.to_pickle(idx2word, save_path+"idx2word.pkl")
    
add_special_token("../tmp/match_w2v_embed.bin", "../tmp/")

Loading [../tmp/match_w2v_embed.bin] embeddding.
Vocabulary Size: 4718
Save add [Pad]、[UNK] Embedding shape: (4720, 100)

任务7、8、11 Common Train Model

任务7的TextCNN、任务8的BiLSTM、任务11的Bert微调我都使用基于Pytorch-lightning框架将模型编码和训练逻辑分开的实现，因此这三个模型共用一个训练模型，实现逻辑如下：

#%%writefile ../src/pl_train_model.py

import torch
import transformers
import pytorch_lightning as pl
torch.set_default_tensor_type(torch.DoubleTensor)
import torch.nn.functional as F

from torchmetrics import Precision
from torchmetrics import Accuracy
from torchmetrics import F1Score
from torchmetrics import AUROC

from model_network import TextCNN
from model_network import BiLSTM
from model_network import Bert
from model_network import roberta_base_AdamW_LLRD

class ClassifyTrainModel(pl.LightningModule):
    def __init__(self, hparams={}):
        
        super().__init__()
        self.save_hyperparameters(hparams)               # 将模型超参数保存在类中，可以通过self.hparams获取
        if self.hparams.encode_model == "TextCNN":
            self.encode = TextCNN(embed_path="../tmp/match_embed.npy")
        elif self.hparams.encode_model == "BiLSTM":
            self.encode = BiLSTM(embed_path="../tmp/match_embed.npy")
        elif self.hparams.encode_model == "Bert":
            self.encode = TransformerEmbed(model_name=self.hparams.pretrain_model,
                                           freeze_layers=self.hparams.freeze_layers,
                                           output_dim=self.hparams.output_dim)
        else:
            print("Error")
        
        self.epochs = self.hparams.epochs
        self.num_steps = self.hparams.num_steps
        self.learning_rate = self.hparams.learning_rate
        
        self.loss= torch.nn.CrossEntropyLoss(weight=torch.DoubleTensor([1.0, 2.0])) #训练集中，正负样本比例接近 1:6
        self.metric_pre = Precision(task="binary", num_classes=self.hparams.output_dim)#.to(device) 
        self.metric_acc = Accuracy(task="binary", num_classes=self.hparams.output_dim)#.to(device)
        self.metric_f1 = F1Score(task="binary", num_classes=self.hparams.output_dim)#.to(device)
        self.metric_auc = AUROC(task="binary", num_classes=self.hparams.output_dim)#.to(device)
    
    
    def forward(self, inputs):
        if self.hparams.encode_model == "TextCNN":
            return self.encode(inputs["input_ids"])
        elif self.hparams.encode_model == "BiLSTM":
            return self.encode(inputs["input_ids"])
        elif self.hparams.encode_model == "Bert":
            return self.encode(inputs["input_ids"], 
                               input["token_type_ids"], 
                               input["attention_mask"])
            
            
    def configure_optimizers(self):
        optimizer = torch.optim.AdamW(filter(lambda p: not p.requires_grad, 
                                             self.parameters()),
                                      lr=self.learning_rate)

        scheduler = transformers.get_scheduler(
            self.hparams.lr_scheduler_type,
            optimizer = optimizer,
            num_warmup_steps = int(self.num_steps*0.1),
            num_training_steps = self.num_steps
        )
        optim_dict = {"optimizer": optimizer, 
                      "lr_scheduler": scheduler}
        return optim_dict
        
        
    def training_step(self, batch, batch_idx):
        logits = self(batch)
        labels = batch["target"]
        if isinstance(labels, list): 
            labels = torch.tensor(labels)
            
        loss = self.loss(logits, labels)
        #loss = torch.nan_to_num(loss)
        self.log("train_step_loss", loss, on_step=True, on_epoch=False, 
                 prog_bar=True, logger=True, sync_dist=True)
        return {"loss": loss}
    
    
    def valid_step(self, batch, batch_idx):
        logits = self(batch)
        labels = batch["target"]
        if isinstance(labels, list):
            labels = torch.tensor(labels)
        
        loss = self.loss(logits, labels)
        loss = torch.nan_to_num(loss)
        preds = torch.argmax(logits, dim=1)
        return {"loss": loss.detach(),
                "preds": preds.detach(),
                "labels": labels}

    
    def valid_epoch_end(self, outputs):
        loss = 0.0
        preds, labels = [], []
        for output in outputs:
            loss += output["loss"]
            preds.append(output["preds"])
            labels.append(output["labels"])

        valid_loss /= len(outputs)
        preds = torch.cat(preds)
        labels = torch.cat(labels)
        valid_pre = self.metric_pre(preds, labels)
        valid_acc = self.metric_acc(preds, labels)
        valid_f1 = self.metric_f1(preds, labels)
        valid_auc = self.metric_auc(preds, labels)

        metrics = {"valid_loss": valid_loss, "valid_pre": valid_pre,
                   "valid_acc": valid_acc, "valid_f1": valid_f1, 
                   "valid_auc": valid_auc}
        self.log_dict(metrics, on_step=False, on_epoch=True,
                      prog_bar=True, logger=True, sync_dist=True)
        return metrics
        
        
    def test_step(self, batch, batch_idx):        
        logits = self(batch)
        labels = batch["target"]
        if isinstance(labels, list):
            labels = torch.tensor(labels)
        
        loss = self.loss(logits, labels)
        preds = torch.argmax(logits, dim=1)
        return {"loss": loss.detach(),
                "preds": preds.detach(),
                "labels": labels}

    
    def test_epoch_end(self, outputs):
        loss = 0.0
        preds, labels = [], []
        for output in outputs:
            loss += output["loss"]
            preds.append(output["preds"])
            labels.append(output["labels"])

        loss /= len(outputs)
        preds = torch.cat(preds)
        labels = torch.cat(labels)
        test_pre = self.metric_pre(preds, labels)
        test_acc = self.metric_acc(preds, labels)
        test_f1 = self.metric_f1(preds, labels)
        test_auc = self.metric_auc(preds, labels)

        metrics = {"test_loss": loss, "test_pre": test_pre,
                   "test_acc": test_acc, "test_f1": test_f1, 
                   "test_auc": test_auc}
        self.log_dict(metrics, on_step=False, on_epoch=True,
                      prog_bar=True, logger=True, sync_dist=True)
        return metrics
    
    
    def predict_step(self, batch, batch_idx):
        logits = self(batch)
        preds = torch.argmax(logits, dim=1)
        return preds.detach()
    
    
    @staticmethod  # 由于该类中并不需要用到实例化对象的属性值，这里使用类的静态方法足以
    def add_model_args(parent_parser):
        parent_parser.add_argument("--output_dim", type=int, default=128)
        parent_parser.add_argument("--num_step", type=int, default=-1)
        parent_parser.add_argument("--learning_rate", type=float, default=0.0004365)
        parent_parser.add_argument("--lr_scheduler_type", type=str, default="cosine")
        parent_parser.add_argument("--freeze_layers", type=list, default=[])
        parent_parser.add_argument("--pretrain_model", type=str, default="hfl/chinese-roberta-wwm-ext")
        return parent_parser

这三个任务，TextCNN和BiLSTM数据输入格式利用预训练好的词向量做一个映射。Bert需要通过transformers库自带tokenizer处理一下。
总共只需要写两种dataset去处理数据，具体逻辑如下：

%%writefile ../src/model_dataset.py

import torch
import numpy as np
import pandas as pd
from torch.utils.data import Dataset

class EmbedDataset(Dataset):
    def __init__(self, vocab_path, df, max_len=512, 
                 pading_index=0, unk_index=1):
        self.vocab = pd.read_pickle(vocab_path)
        self.df = df
        self.comment_text = self.df.content
        self.max_len = max_len
        if "label" in self.df:
            self.targets = self.df['label'].tolist()
    
    def __len__(self):
        return len(self.comment_text)

    def get_id(self, word):
        if word not in self.vocab.keys():
            return 1
        else:
            return self.vocab[word]
        
    def __getitem__(self, index):
        text = self.comment_text[index].split(" ")
        text = text[:self.max_len]
        input_ids = list(map(lambda x: self.get_id(x), text))
        if len(input_ids) < self.max_len:
            input_ids += [0] * (self.max_len - len(input_ids))  # padding 0 
        ans = {"input_ids": input_ids}
        if "label" in self.df:
            ans["target"] = self.targets[index]
        return ans


class BertDataset(Dataset):
    def __init__(self, df, tokenizer, max_len=512):
        self.tokenizer = tokenizer
        self.df = df
        self.comment_text = self.df.content
        self.max_len = max_len
        if "label" in self.df:
            self.targets = self.df['label'].tolist()

    def __len__(self):
        return len(self.comment_text)

    def __getitem__(self, index):
        comment_text = self.comment_text[index]
        encode_dict = self.tokenizer.encode_plus(
            comment_text,
            add_special_tokens=True,     # add token [CLS], [SEP]
            truncation=True,             # whether to truncate long sentences
            max_length=self.max_len,     # Maximum sentence length
            padding="max_length",        # Insufficient length Fill to maximum lengt
            return_attention_mask=True,  # Whether to return attention mask
            return_tensors="pt",         # Return tensor type
        )
        input_ids = encode_dict["input_ids"]
        token_type_ids = encode_dict["token_type_ids"]
        attention_mask = encode_dict['attention_mask']
        
        ans = {"input_ids": input_ids.squeeze(0),                # (1, max_len) -> (max_len)
               "token_type_ids": token_type_ids.squeeze(0),      # (1, max_len) -> (max_len)
               "attention_mask": attention_mask.squeeze(0)}      # (1, max_len) -> (max_len)
        
        if "label" in self.df:
            ans["target"] = self.targets[index]
            
        return ans

Overwriting ../src/model_dataset.py

任务七：使用Word2Vec词向量，搭建TextCNN模型进行训练和预测

上面已经写好训练网络的框架和处理数据的逻辑，下面都只需要关注网络结构的编码，TextCNN的网络结构编码如下：

import numpy as np
import torch
import torch.nn.functional as F
from torch import nn


class Mish(nn.Module):  # 定义mish激活函数
    def __init__(self):
        super(Mish, self).__init__()
        
    def forward(self,x):
        x = x * (torch.tanh(F.softplus(x)))
        return x

class TextCNN(nn.Module):
    def __init__(self, embed_path, n_filters=100,
                 filter_sizes=[2, 3, 4], output_dim=2, dropout=0.1, pad_idx=0):

        super().__init__()
        embedding = np.load(embed_path)
        vocab_size, embedding_dim = embedding.shape
        self.embedding = nn.Embedding(vocab_size, embedding_dim, 
                                      padding_idx=pad_idx)
        
        self.embedding.weight.data.copy_(torch.from_numpy(embedding))
        self.embedding.weight.requires_grad = False
        self.convs = nn.ModuleList([nn.Conv2d(in_channels=1,
                                              out_channels=n_filters,
                                              kernel_size=(fs, embedding_dim))
                                    for fs in filter_sizes])
        
        self.fc = nn.Linear(len(filter_sizes) * n_filters, output_dim)
        self.dropout = nn.Dropout(dropout)
        self.mish = Mish()
 
    def forward(self, inputs):
        embedded = self.embedding(inputs)   # [batch size, sen len, embed dim]
        print(embedded.shape)
        embedded = embedded.unsqueeze(1)  # [batch size, 1, sen len, embed dim]
        print(embedded.shape)
        conved = [self.mish(conv(embedded)).squeeze(3) for conv in self.convs]
        pooled = [F.max_pool1d(conv, conv.shape[2]).squeeze(2) for conv in conved]
    
        cat = self.dropout(torch.cat(pooled, dim=1))  # [batch size, n_filters * len(filter_sizes)]
        return self.fc(cat)

text_cnn = TextCNN(embed_path="../tmp/match_embed.npy")

TextCNN 训练的过程如下：

import argparse
import torch
from torch.utils.data import DataLoader
device = "cuda" if torch.cuda.is_available() else "cpu"
torch.set_default_tensor_type(torch.DoubleTensor)

import pytorch_lightning as pl

from sklearn.model_selection import train_test_split
from pytorch_lightning import Trainer
from pytorch_lightning.loggers import TensorBoardLogger
from sklearn.model_selection import train_test_split

from model_dataset import EmbedDataset
from pl_train_model import ClassifyTrainModel

hparams = {'encode_model': 'textcnn',
           'use_swa': True,
           'max_len': 512,
           'output_dim': 2,
           'num_works': 10,
           'batch_size': 8,
           'accumulate_grad_batches': 16,
           'pin_memory': True,
           'epochs': 10,
           'num_steps': -1,
           'learning_rate': 1e-5,
           'auto_lr_find': False,
           'lr_scheduler_type': 'adaw',
           'freeze_layers': [],
           'precision': 16,
           'train_monitor': 'valid_f1',
           'train_mode': 'max',
           'save_dir': './models/textcnn_models/',
           'log_dir': './logs/',
           'accelerator': 'gpu',
           'devices': 1}

hparams = argparse.Namespace(**hparams)

train,  _ = load_data()
train = train[1:]
train['label'] = train['label'].apply(lambda x: int(x))
train_more, test_more = train_test_split(train, 
                                         test_size=0.1, 
                                         random_state=43)
train_more = train_more.reset_index()
test_more = test_more.reset_index()
train_dataset = EmbedDataset(train_more, tokenizer, max_len=hparams.max_len)
train_dataloader = DataLoader(train_dataset, 
                              shuffle=True,
                              num_workers=hparams.num_works,
                              batch_size=hparams.batch_size,
                              pin_memory=True)
hparams.num_steps = len(train_dataloader) * hparams.epochs
test_dataset =  EmbedDataset(test_more, tokenizer, max_len=hparams.max_len)
test_dataloader = DataLoader(test_dataset, 
                             shuffle=True,
                             num_workers=hparams.num_works,
                             batch_size=hparams.batch_size,
                             pin_memory=True)

pl.seed_everything(1234)  # 统一设置随机种子
train_model = ClassifyTrainModel(hparams)
ckpt_callback = pl.callbacks.ModelCheckpoint(
    monitor=hparams.train_monitor,
    dirpath=hparams.save_dir,
    filename="%s-{epochs:03d}-{%s:.3f}" % (hparams.encode_model, 
                                           hparams.train_monitor),
    mode=hparams.train_mode)

callbacks = [ckpt_callback]
callbacks.append(pl.callbacks.StochasticWeightAveraging(swa_lrs=0.05))

logger = TensorBoardLogger(save_dir=hparams.log_dir, name="TrainModel")
trainer = pl.Trainer.from_argparse_args(
    hparams,
    max_epochs=hparams.epochs,
    logger=logger,
    callbacks=callbacks)

print("hparams.auto_lr_find=", hparams.auto_lr_find)
if hparams.auto_lr_find:
    lr_finder = trainer.tuner.lr_find(model=train_model,
         train_dataloaders=train_dataloader,
         min_lr=1e-08,
         max_lr=1e-1,
         num_training=100,
         mode="exponential",              # 寻找策略:指数型exponential 和线数型 linear
         early_stop_threshold=4.0)

    fig = lr_finder.plot(suggest=True)
    fig.show()
    lr = lr_finder.suggestion()
    print("suggest lr=", lr)
    del train_model
    hparams.learning_rate = lr
    train_model = BertTrainModel(hparams)
    print("Search best learning_rate: [%f]" % train_model.learning_rate)

trainer.fit(train_model, train_dataloader)
print("Test Data Result:")
result = trainer.test(train_model, test_dataloader)
print("Result:", result)

任务八：使用Word2Vec词向量，搭建BILSTM模型进行训练和预测

import numpy as np
import torch
import torch.nn.functional as F
from torch import nn

class BiLSTM(nn.Module):
    def __init__(self, embed_path, n_hidden=100, 
                 output_dim=2, num_layers=2, dropout=0.1, pad_idx=0):

        super().__init__()
        embedding = np.load(embed_path)
        vocab_size, embedding_dim = embedding.shape
        self.embedding = nn.Embedding(vocab_size, embedding_dim, 
                                      padding_idx=pad_idx)
        
        self.embedding.weight.data.copy_(torch.from_numpy(embedding))
        self.embedding.weight.requires_grad = False
        self.bilstm = nn.LSTM(input_size=embedding_dim, hidden_size=n_hidden, 
                              bidirectional=True, batch_first=True, num_layers=num_layers)
        self.dropout = nn.Dropout(dropout)
        self.fc = nn.Linear(2*n_hidden, output_dim)
 
    def forward(self, inputs):
        embed = self.embedding(inputs)                    # [batch size, sen len, embed dim]
        print(embed.shape)
        _, (h_n, _) = self.bilstm(embed)                  # bilstm output : [batch_size, seq_len, hidden_dim * 2]
        output = torch.cat((h_n[-1], h_n[-2]), dim=-1)
        dropout = self.dropout(output)                    
        return self.fc(dropout)

bilstm = BiLSTM(embed_path="../tmp/match_embed.npy")
bilstm.embedding

import argparse
import torch
from torch.utils.data import DataLoader
device = "cuda" if torch.cuda.is_available() else "cpu"
torch.set_default_tensor_type(torch.DoubleTensor)

import pytorch_lightning as pl

from sklearn.model_selection import train_test_split
from pytorch_lightning import Trainer
from pytorch_lightning.loggers import TensorBoardLogger
from sklearn.model_selection import train_test_split

from model_dataset import EmbedDataset
from pl_train_model import ClassifyTrainModel

hparams = {'encode_model': 'bilstm',
           'use_swa': True,
           'max_len': 512,
           'output_dim': 2,
           'num_works': 10,
           'batch_size': 8,
           'accumulate_grad_batches': 16,
           'pin_memory': True,
           'epochs': 10,
           'num_steps': -1,
           'learning_rate': 1e-5,
           'auto_lr_find': False,
           'lr_scheduler_type': 'adaw',
           'freeze_layers': [],
           'precision': 16,
           'train_monitor': 'valid_f1',
           'train_mode': 'max',
           'save_dir': './models/textcnn_models/',
           'log_dir': './logs/',
           'accelerator': 'gpu',
           'devices': 1}

hparams = argparse.Namespace(**hparams)

train,  _ = load_data()
train = train[1:]
train['label'] = train['label'].apply(lambda x: int(x))
train_more, test_more = train_test_split(train, 
                                         test_size=0.1, 
                                         random_state=43)
train_more = train_more.reset_index()
test_more = test_more.reset_index()
train_dataset = EmbedDataset(train_more, tokenizer, max_len=hparams.max_len)
train_dataloader = DataLoader(train_dataset, 
                              shuffle=True,
                              num_workers=hparams.num_works,
                              batch_size=hparams.batch_size,
                              pin_memory=True)
hparams.num_steps = len(train_dataloader) * hparams.epochs
test_dataset =  EmbedDataset(test_more, tokenizer, max_len=hparams.max_len)
test_dataloader = DataLoader(test_dataset, 
                             shuffle=True,
                             num_workers=hparams.num_works,
                             batch_size=hparams.batch_size,
                             pin_memory=True)

pl.seed_everything(1234)  # 统一设置随机种子
train_model = ClassifyTrainModel(hparams)
ckpt_callback = pl.callbacks.ModelCheckpoint(
    monitor=hparams.train_monitor,
    dirpath=hparams.save_dir,
    filename="%s-{epochs:03d}-{%s:.3f}" % (hparams.encode_model, 
                                           hparams.train_monitor),
    mode=hparams.train_mode)

callbacks = [ckpt_callback]
callbacks.append(pl.callbacks.StochasticWeightAveraging(swa_lrs=0.05))

logger = TensorBoardLogger(save_dir=hparams.log_dir, name="TrainModel")
trainer = pl.Trainer.from_argparse_args(
    hparams,
    max_epochs=hparams.epochs,
    logger=logger,
    callbacks=callbacks)

print("hparams.auto_lr_find=", hparams.auto_lr_find)
if hparams.auto_lr_find:
    lr_finder = trainer.tuner.lr_find(model=train_model,
         train_dataloaders=train_dataloader,
         min_lr=1e-08,
         max_lr=1e-1,
         num_training=100,
         mode="exponential",              # 寻找策略:指数型exponential 和线数型 linear
         early_stop_threshold=4.0)

    fig = lr_finder.plot(suggest=True)
    fig.show()
    lr = lr_finder.suggestion()
    print("suggest lr=", lr)
    del train_model
    hparams.learning_rate = lr
    train_model = BertTrainModel(hparams)
    print("Search best learning_rate: [%f]" % train_model.learning_rate)

trainer.fit(train_model, train_dataloader)
print("Test Data Result:")
result = trainer.test(train_model, test_dataloader)
print("Result:", result)

任务九：学会Bert基础，transformer库基础使用

from transformers import BertConfig
from transformers import BertTokenizer
from transformers import BertModel

# 使用huggingface transformers库下载 hfl/chinese-macbert-base 的 Bert模型到本地
model_name = "hfl/chinese-macbert-base"        
model_save = "./pretrain_models/hfl/chinese-macbert-base/"  


# 从本地 cache_dir 载入，初始化模型
config = BertConfig.from_pretrained(model_name, cache_dir=model_save)
config.output_hidden_states = True                                       # 
tokenizer = BertTokenizer.from_pretrained(model_name, cache_dir=model_save)
model = BertModel.from_pretrained(model_name, cache_dir=model_save)


# 输入文本，产生编码
sens = ["Hugging Face是什么，提供了哪些内容？", 
         "近年来人工智能在自然语言处理领域取得了巨大的进展。其中一项引人注目的技术是生成模型，如OpenAI的GPT-3.5。"]

print("tokenizer.tokenize：", tokenizer.tokenize(sens[0]))

input_ids = tokenizer(sens,
                      add_special_tokens=True, 
                      return_attention_mask=True,
                      return_tensors="pt",
                      padding=True,
                      truncation="longest_first",
                      max_length=128)

print("input ids:", input_ids)
print("input ids shape:", input_ids["input_ids"].shape)


# 利用bert模型输出进行特征提取

# hidden_states 包含bert输入层embedding和每一层输出的embedding, 维度都是[batch, seq_len, hidden_state]
hidden_states = model(input_ids["input_ids"], 
                      input_ids["token_type_ids"], 
                      input_ids["attention_mask"], 
                      output_hidden_states=True).hidden_states

# bert常见的不同特征提取方式
def bert_pooling(hidden_states, pooling="last_cls"):
    if pooling == "last_cls":
        return hidden_states[-1][:, 0, :]
    elif pooling == "first_last_avg":
        return (hidden_states[1] + hidden_states[-1]).mean(dim=1)
    elif pooling == "last_avg":
        return (hidden_states[-1]).mean(dim=1)
    elif pooling == 'last2avg':
        return (hidden_states[-1] + hidden_states[-2]).mean(dim=1)
    else:
        raise Exception("unknown pooling {}".format(self.pooling))

# 使用输出的第一层和最后一层相加+平均池化的方式
output = bert_pooling(hidden, pooling="first_last_avg")
print("bert output shape:", output.shape)   # [2, 768]

任务十：使用Bert在比赛数据集中完成预训练

生成预训练数据

# 将比赛数据生成预训练语料

all_text = []
token = set()

for data in train["content"].tolist():
    for line in data[1:-1].split("\n "):
        all_text.append(line)
        for word in line.split(" "):
            token.add(word)

for data in test["content"].tolist():
    for line in data[1:-1].split("\n "):
        all_text.append(line)
        for word in line.split(" "):
            token.add(word)
  
with open("./data/all_data_txt.txt", "w") as file:
    for line in all_text:
        file.write(line + "\n")

len(token)

重新制作 Tokenizer，生成vocab.txt

# code encourage by https://blog.csdn.net/qq_26593695/article/details/115338593

from transformers import BertTokenizer

import tokenizers
# 创建分词器
bwpt = tokenizers.BertWordPieceTokenizer()
filepath = "./data/all_data_txt.txt" # 语料文件
#训练分词器
bwpt.train(
    files=[filepath],
    vocab_size=len(token), # 这里预设定的词语大小不是很重要
    min_frequency=1,
    limit_alphabet=1000
)
# 保存训练后的模型词表
bwpt.save_model('./models/pretrain_models/')
#output： ['./data/pre_train/vocab.txt']

# 加载刚刚训练的tokenizer
tokenizer = BertTokenizer(vocab_file='./models/pretrain_models/vocab.txt')

from transformers import (
    CONFIG_MAPPING,
    MODEL_FOR_MASKED_LM_MAPPING, 
    AutoConfig,
    AutoModelForMaskedLM,
    AutoTokenizer,
    DataCollatorForLanguageModeling,
    HfArgumentParser,
    Trainer,
    TrainingArguments,
    set_seed,
    LineByLineTextDataset)

# 自己修改部分配置参数
config_kwargs = {
    "cache_dir": None,
    "revision": 'main',
    "use_auth_token": None,
     "hidden_size": 512,
    "num_attention_heads": 4,
    "hidden_dropout_prob": 0.2,
    "vocab_size": len(token) # 自己设置词汇大小
}

# 将模型的配置参数载入
#config = AutoConfig.from_pretrained('', **config_kwargs)
# 载入预训练模型
model_name = "sijunhe/nezha-cn-base"
model_save = "../models/pretrain_models/sijunhe/nezha-cn-base/"
model = AutoModelForMaskedLM.from_pretrained(
            model_name,
            cache_dir=model_save, 
            revision='main'
        )
model.resize_token_embeddings(len(tokenizer))
#output:Embedding(5230, 768, padding_idx=1)

# 通过LineByLineTextDataset接口 加载数据 #长度设置为512, # 这里file_path于本文第一部分的语料格式一致
train_dataset = LineByLineTextDataset(tokenizer=tokenizer,
                                      file_path=filepath,
                                      block_size=512)
# MLM模型的数据DataCollator
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, 
                                                mlm=True, 
                                                mlm_probability=0.15)

# 训练参数
pretrain_batch_size = 128
num_train_epochs = 100
training_args = TrainingArguments(
    output_dir='./data/pre_train/',
    resume_from_checkpoint=True,
    overwrite_output_dir=True,
    num_train_epochs=num_train_epochs,
    learning_rate=8e-5,
    lr_scheduler_type="cosine",
    warmup_ratio=0.1,
    weight_decay=0.01,
    per_device_train_batch_size=pretrain_batch_size,
    gradient_accumulation_steps=8,
    logging_steps=100,
    save_strategy='steps', 
    save_steps=1000,       # 每1000 steps保存一个模型
    save_total_limit=10)   # 保存最近10个checkpoint,控制保存checkpoint的数量

# 通过Trainer接口训练模型
trainer = Trainer(
    model=model,
    args=training_args,
    data_collator=data_collator,
    train_dataset=train_dataset)

# 开始训练
trainer.train(True)    # 之前有保存的模型
#trainer.train()       # 第一次训练
trainer.save_model('./data/pre_train/')

Some weights of the model checkpoint at sijunhe/nezha-cn-base were not used when initializing NezhaForMaskedLM: ['cls.seq_relationship.bias', 'cls.seq_relationship.weight']
- This IS expected if you are initializing NezhaForMaskedLM 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 NezhaForMaskedLM from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
Embedding(5230, 768)

任务十一：使用Bert在比赛数据集上完成微调

Model Network

%%writefile ../src/model_network.py

import os
import torch
import torch.nn.functional as F
from torch import nn

import transformers
from transformers import BertModel

class Bert(nn.Module):
    def __init__(self, model_name="hfl/chinese-roberta-wwm-ext/",
                 pooling="last_avg",
                 whitening=False,
                 output_dim=128,
                 hidden_size=768,
                 model_dir="../models/pretrain_models/",
                 freeze_layers=[]):
        
        super().__init__()
        self.model_name = model_name
        if os.path.exists(model_dir+model_name):
            model_name = model_dir + model_name
        self.sen_bert = BertModel.from_pretrained(model_name)
        self.freeze_layers = freeze_layers
        if len(self.freeze_layers) > 0:
            self.freeze_params()
        self.pooling = pooling
        self.Linear = nn.Linear(hidden_size, output_dim)
        self.whitening = whitening
    
    def freeze_params(self):
        for name, param in self.sen_bert.named_parameters():
            for freeze_name in self.freeze_layers:
                if freeze_name in name:
                    param.requires_grad = False
                    break
    
    def bert_pooling(self, hidden_states):
        if self.pooling == "last_cls":
            return hidden_states[-1][:, 0, :]
        elif self.pooling == "first_last_avg":
            return (hidden_states[1] + hidden_states[-1]).mean(dim=1)
        elif self.pooling == "last_avg":
            return (hidden_states[-1]).mean(dim=1) 
        elif self.pooling == 'last2avg':
            return (hidden_states[-1] + hidden_states[-2]).mean(dim=1)
        else:
            raise Exception("unknown pooling {}".format(self.pooling))
    
    def bert_whitening(self, output): # 暂时不折腾
        pass
    
    def forward(self, input_ids, token_type_ids, attention_mask):
        hidden_states = self.sen_bert(input_ids,
                                      token_type_ids,
                                      attention_mask,
                                      output_hidden_states=True).hidden_states
        if self.whitening:
            pass
            #return self.bert_whitening(output)
        else:
            output_embed = self.bert_pooling(hidden_states)  # [batch, hidden_size]
        
        output_embed = self.Linear(output_embed) # [batch, output_dim]
        return output_embed

Overwriting ../src/model_network.py

Bert微调的逻辑：

import argparse
import torch
from torch.utils.data import DataLoader
device = "cuda" if torch.cuda.is_available() else "cpu"
torch.set_default_tensor_type(torch.DoubleTensor)

import tokenizers
from transformers import BertTokenizer
tokenizer = BertTokenizer(vocab_file='../models/pretrain_models/vocab.txt')

import pytorch_lightning as pl

from sklearn.model_selection import train_test_split
from pytorch_lightning import Trainer
from pytorch_lightning.loggers import TensorBoardLogger
from sklearn.model_selection import train_test_split

from model_dataset import CustomDataset
from pl_train_model import BertTrainModel

hparams = {'use_swa': True,
           'max_len': 512,
           'output_dim': 2,
           'num_works': 10,
           'batch_size': 8,
           'accumulate_grad_batches': 16,
           'pin_memory': True,
           'epochs': 10,
           'num_steps': -1,
           'pretrain_model': 'checkpoint-35000',
           'learning_rate': 1e-5,
           'auto_lr_find': False,
           'lr_scheduler_type': 'cosine',
           'freeze_layers': [],
           'precision': 16,
           'train_monitor': 'train_step_loss',
           'train_mode': 'min',
           'save_dir': './models/finetune_models/',
           'log_dir': './logs/',
           'accelerator': 'gpu',
           'devices': 1}

hparams = argparse.Namespace(**hparams)

train,  _ = load_data()
train = train[1:]
train['label'] = train['label'].apply(lambda x: int(x))
train_more, test_more = train_test_split(train, 
                                         test_size=0.1, 
                                         random_state=43)
train_more = train_more.reset_index()
test_more = test_more.reset_index()
train_dataset = CustomDataset(train_more, tokenizer, max_len=hparams.max_len)
train_dataloader = DataLoader(train_dataset, 
                              shuffle=True,
                              num_workers=hparams.num_works,
                              batch_size=hparams.batch_size,
                              pin_memory=True)
hparams.num_steps = len(train_dataloader) * hparams.epochs
test_dataset =  CustomDataset(test_more, tokenizer, max_len=hparams.max_len)
test_dataloader = DataLoader(test_dataset, 
                             shuffle=True,
                             num_workers=hparams.num_works,
                             batch_size=hparams.batch_size,
                             pin_memory=True)

pl.seed_everything(1234)  # 统一设置随机种子


train_model = ClassifyTrainModel(hparams)
ckpt_callback = pl.callbacks.ModelCheckpoint(
    monitor=hparams.train_monitor,
    dirpath=hparams.save_dir,
    filename="finetune-{epochs:03d}-{%s:.3f}" % hparams.train_monitor,
    mode=hparams.train_mode)

callbacks = [ckpt_callback]
callbacks.append(pl.callbacks.StochasticWeightAveraging(swa_lrs=0.05))

logger = TensorBoardLogger(save_dir=hparams.log_dir, name="TrainModel")
trainer = pl.Trainer.from_argparse_args(
    hparams,
    max_epochs=hparams.epochs,
    logger=logger,
    callbacks=callbacks)

print("hparams.auto_lr_find=", hparams.auto_lr_find)
if hparams.auto_lr_find:
    lr_finder = trainer.tuner.lr_find(model=train_model,
         train_dataloaders=train_dataloader,
         min_lr=1e-08,
         max_lr=1e-1,
         num_training=100,
         mode="exponential",              # 寻找策略:指数型exponential 和线数型 linear
         early_stop_threshold=4.0)

    fig = lr_finder.plot(suggest=True)
    fig.show()
    lr = lr_finder.suggestion()
    print("suggest lr=", lr)
    del train_model
    hparams.learning_rate = lr
    train_model = BertTrainModel(hparams)
    # 等价于
    # trainer.tune(train_model, data_module)
    print("Search best learning_rate: [%f]" % train_model.learning_rate)

trainer.fit(train_model, train_dataloader)
print("Test Data Result:")
result, preds, labels = trainer.test(train_model, test_dataloader)
print("Result:", result)