2 文本分类入门：TextCNN

论文链接：https://arxiv.org/pdf/1408.5882.pdf

TextCNN 是一种用于文本分类的卷积神经网络模型。它在卷积神经网络的基础上进行了一些修改，以适应文本数据的特点。

TextCNN 的主要思想是使用一维卷积层来提取文本中的局部特征，并通过池化操作来减少特征的维度。这些局部特征可以捕获词语之间的关系和重要性，从而帮助模型进行分类。

nn.Conv2d

nn.Conv2d 的构造函数包含以下参数：

• in_channels：输入数据的通道数。
• out_channels：卷积核的数量，也是输出数据的通道数。
• kernel_size：卷积核的大小，可以是一个整数或一个元组，表示宽度和高度。
• stride：卷积核的步幅，可以是一个整数或一个元组，表示水平和垂直方向的步幅。

nn.Conv2d(1, config.num_filters, (k, config.embed))

输入通道是1 ， 输出通道的维度， 卷积核（k, config.embed)）

代码部分：

import pandas as pd
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import pickle as pkl
from tqdm import tqdm
import time
from torch.utils.data import Dataset

from datetime import timedelta

from sklearn.model_selection import train_test_split
from torch.utils.data import Dataset, DataLoader
from collections import defaultdict
from torch.optim import AdamW


df = pd.read_csv("./data/online_shopping_10_cats.csv")
UNK, PAD = '<UNK>', '<PAD>'  # 未知字，padding符号
RANDOM_SEED = 2023


file_path = "./data/online_shopping_10_cats.csv"
vocab_file = "./data/vocab.pkl"
emdedding_file = "./data/embedding_SougouNews.npz"
vocab = pkl.load(open(vocab_file, 'rb'))

class MyDataSet(Dataset):
    def __init__(self, df, vocab,pad_size=None):
        self.data_info = df
        self.data_info['review'] = self.data_info['review'].apply(lambda x:str(x).strip())
        self.data_info = self.data_info[['review','label']].values
        self.vocab = vocab 
        self.pad_size = pad_size
        self.buckets = 250499  
        
    def biGramHash(self,sequence, t):
        t1 = sequence[t - 1] if t - 1 >= 0 else 0
        return (t1 * 14918087) % self.buckets
        
    def triGramHash(self,sequence, t):
        t1 = sequence[t - 1] if t - 1 >= 0 else 0
        t2 = sequence[t - 2] if t - 2 >= 0 else 0
        return (t2 * 14918087 * 18408749 + t1 * 14918087) % self.buckets
        
    def __getitem__(self, item):
        result = {}
        view, label = self.data_info[item]
        result['view'] = view.strip()
        result['label'] = torch.tensor(label,dtype=torch.long)
        
        token = [i for i in view.strip()]
        seq_len = len(token)
        # 填充
        if self.pad_size:
            if len(token) < self.pad_size:
                token.extend([PAD] * (self.pad_size - len(token)))
            else:
                token = token[:self.pad_size]
                seq_len = self.pad_size
        result['seq_len'] = seq_len
        
        # 词表的转换
        words_line = []
        for word in token:
            words_line.append(self.vocab.get(word, self.vocab.get(UNK)))
        result['input_ids'] = torch.tensor(words_line, dtype=torch.long) 
        
        # 
        bigram = []
        trigram = []
        for i in range(self.pad_size):
            bigram.append(self.biGramHash(words_line, i))
            trigram.append(self.triGramHash(words_line, i))
            
        result['bigram'] = torch.tensor(bigram, dtype=torch.long)
        result['trigram'] = torch.tensor(trigram, dtype=torch.long)
        return result

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


#myDataset[0]
df_train, df_test = train_test_split(df, test_size=0.1, random_state=RANDOM_SEED)
df_val, df_test = train_test_split(df_test, test_size=0.5, random_state=RANDOM_SEED)
df_train.shape, df_val.shape, df_test.shape

#((56496, 3), (3139, 3), (3139, 3))


def create_data_loader(df,vocab,pad_size,batch_size=4):
    ds = MyDataSet(df,
                   vocab,
                   pad_size=pad_size
                  )
    return DataLoader(ds,batch_size=batch_size)

MAX_LEN = 256
BATCH_SIZE = 4
train_data_loader = create_data_loader(df_train,vocab,pad_size=MAX_LEN, batch_size=BATCH_SIZE)
val_data_loader = create_data_loader(df_val,vocab,pad_size=MAX_LEN, batch_size=BATCH_SIZE)
test_data_loader = create_data_loader(df_test,vocab,pad_size=MAX_LEN, batch_size=BATCH_SIZE)

class Config(object):

    """配置参数"""
    def __init__(self):
        self.model_name = 'FastText'
        self.embedding_pretrained = torch.tensor(
            np.load("./data/embedding_SougouNews.npz")["embeddings"].astype('float32'))  # 预训练词向量
            
        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')   # 设备

        self.dropout = 0.5                                              # 随机失活
        self.require_improvement = 1000                                 # 若超过1000batch效果还没提升，则提前结束训练
        self.num_classes = 2                                            # 类别数
        self.n_vocab = 0                                                # 词表大小，在运行时赋值
        self.num_epochs = 20                                            # epoch数
        self.batch_size = 128                                           # mini-batch大小
        self.learning_rate = 1e-4                                       # 学习率
        self.embed = self.embedding_pretrained.size(1)\
            if self.embedding_pretrained is not None else 300           # 字向量维度
        self.hidden_size = 256                                          # 隐藏层大小
        self.n_gram_vocab = 250499                                      # ngram 词表大小
        self.filter_sizes = [2,3,4]
        self.num_filters = 256                                          # 卷积核数量(channels数)


class Model(nn.Module):
    def __init__(self, config):
        super(Model, self).__init__()
        if config.embedding_pretrained is not None:
            self.embedding = nn.Embedding.from_pretrained(config.embedding_pretrained, freeze=False)
        else:
            self.embedding = nn.Embedding(config.n_vocab, config.embed, padding_idx=config.n_vocab - 1)
        self.convs = nn.ModuleList(
            [nn.Conv2d(1, config.num_filters, (k, config.embed)) for k in config.filter_sizes])

        # self.convs = nn.ModuleList(
        #     [nn.Conv1D(1, config.num_filters, k) for k in config.filter_sizes]
        # )
        self.dropout = nn.Dropout(config.dropout)
        self.fc = nn.Linear(config.num_filters * len(config.filter_sizes), config.num_classes)

    def conv_and_pool(self, x, conv):
        x = F.relu(conv(x)).squeeze(3)
        x = F.max_pool1d(x, x.size(2)).squeeze(2)
        return x

    def forward(self, x):
        out = self.embedding(x['input_ids'])
        out = out.unsqueeze(1)
        out = torch.cat([self.conv_and_pool(out, conv) for conv in self.convs], 1)
        out = self.dropout(out)
        out = self.fc(out)
        return out


config = Config()
model = Model(config)
sample = next(iter(train_data_loader))


device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = model.to(device)
 
EPOCHS = 5 # 训练轮数
optimizer = AdamW(model.parameters(),lr=2e-4)
total_steps = len(train_data_loader) * EPOCHS
# schedule = get_linear_schedule_with_warmup(optimizer,num_warmup_steps=0,
#                                num_training_steps=total_steps)
loss_fn = nn.CrossEntropyLoss().to(device)

def train_epoch(model,data_loader,loss_fn,device, optimizer,n_examples,schedule=None):
    model = model.train()
    losses = []
    correct_predictions = 0
    for d in tqdm(data_loader):
        # input_ids = d['input_ids'].to(device)
        # attention_mask = d['attention_mask'].to(device)
        targets = d['label']#.to(device)
        outputs = model(d)
        
        _,preds = torch.max(outputs, dim=1)
        loss = loss_fn(outputs,targets)
        losses.append(loss.item())
        
        correct_predictions += torch.sum(preds==targets)
        loss.backward()
        nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
        optimizer.step()
        #scheduler.step()
        optimizer.zero_grad()
        #break
    #print(n_examples)
    return correct_predictions.double().item() / n_examples, np.mean(losses)
 
def eval_model(model, data_loader, loss_fn, device, n_examples):
    model = model.eval() # 验证预测模式
 
    losses = []
    correct_predictions = 0
 
    with torch.no_grad():
        for d in data_loader:
            targets = d['label']#.to(device)
            outputs = model(d)
            _, preds = torch.max(outputs, dim=1)
 
            loss = loss_fn(outputs, targets)
 
            correct_predictions += torch.sum(preds == targets)
            losses.append(loss.item())
        
    return correct_predictions.double() / n_examples, np.mean(losses)

# train model
EPOCHS = 10
history = defaultdict(list) # 记录10轮loss和acc
best_accuracy = 0
 
for epoch in range(EPOCHS):
 
    print(f'Epoch {epoch + 1}/{EPOCHS}')
    print('-' * 10)
 
    train_acc, train_loss = train_epoch(
        model,
        train_data_loader,
        loss_fn = loss_fn,
        optimizer=optimizer,
        device = device,
        n_examples=len(df_train)
    )
 
    print(f'Train loss {train_loss} accuracy {train_acc}')
 
    val_acc, val_loss = eval_model(
        model,
        val_data_loader,
        loss_fn,
        device,
        len(df_val)
    )
 
    print(f'Val   loss {val_loss} accuracy {val_acc}')
    print()
 
    history['train_acc'].append(train_acc)
    history['train_loss'].append(train_loss)
    history['val_acc'].append(val_acc)
    history['val_loss'].append(val_loss)
 
    if val_acc > best_accuracy:
        torch.save(model.state_dict(), 'best_model_state.bin')
        best_accuracy = val_acc

一维卷积模型,直接替换就行了

class Model(nn.Module):
    def __init__(self, config):
        super(Model, self).__init__()
        if config.embedding_pretrained is not None:
            self.embedding = nn.Embedding.from_pretrained(config.embedding_pretrained, freeze=False)
        else:
            self.embedding = nn.Embedding(config.n_vocab, config.embed, padding_idx=config.n_vocab - 1)
        # self.convs = nn.ModuleList(
        #     [nn.Conv2d(1, config.num_filters, (k, config.embed)) for k in config.filter_sizes])

        self.convs = nn.ModuleList(
            [nn.Conv1d(MAX_LEN, config.num_filters, k) for k in config.filter_sizes]
        )
        self.dropout = nn.Dropout(config.dropout)
        self.fc = nn.Linear(config.num_filters * len(config.filter_sizes), config.num_classes)

    def conv_and_pool(self, x, conv):
        #print(x.shape)
        x = F.relu(conv(x))#.squeeze(3)
        #print(x.shape)
        x = F.max_pool1d(x, x.size(2))#.squeeze(2)
        return x

    def forward(self, x):
        out = self.embedding(x['input_ids'])
        #print(out.shape)
        #out = out.unsqueeze(1)
        out = torch.cat([self.conv_and_pool(out, conv) for conv in self.convs], 1)
        out = out.squeeze(-1)
        #print(out.shape)
        out = self.fc(out)
        return out

Epoch 1/10
----------

100%|█████████████████████████████████████| 14124/14124 [08:19<00:00, 28.29it/s]

Train loss 0.32963800023092527 accuracy 0.889903709997168
Val   loss 0.2872631916414839 accuracy 0.9197196559413826

Epoch 2/10
----------

100%|█████████████████████████████████████| 14124/14124 [08:19<00:00, 28.25it/s]

Train loss 0.26778308933985917 accuracy 0.925392948173322
Val   loss 0.29051536209677714 accuracy 0.9238611022618668

Epoch 3/10
----------

100%|█████████████████████████████████████| 14124/14124 [08:17<00:00, 28.39it/s]

Train loss 0.23998896145841375 accuracy 0.9368450863777966
Val   loss 0.29530937147389363 accuracy 0.9238611022618668

Epoch 4/10
----------

100%|█████████████████████████████████████| 14124/14124 [08:21<00:00, 28.14it/s]

Train loss 0.21924698638110582 accuracy 0.9446863494760691
Val   loss 0.3079132618505083 accuracy 0.9260911118190507

Epoch 5/10
----------

100%|█████████████████████████████████████| 14124/14124 [08:21<00:00, 28.15it/s]

Train loss 0.1976975509786261 accuracy 0.9515717926932881
Val   loss 0.3294101043627459 accuracy 0.9267282574068174

Epoch 6/10
----------

100%|█████████████████████████████████████| 14124/14124 [08:14<00:00, 28.56it/s]

Train loss 0.18130036814091913 accuracy 0.9575899178702917
Val   loss 0.34197808585767564 accuracy 0.9260911118190507

Epoch 7/10
----------

100%|█████████████████████████████████████| 14124/14124 [09:03<00:00, 26.00it/s]

Train loss 0.16165128718584662 accuracy 0.9624044180118947
Val   loss 0.34806641904714486 accuracy 0.924816820643517

conv1D：

Epoch 1/10
----------

100%|█████████████████████████████████████| 14124/14124 [04:53<00:00, 48.14it/s]

Train loss 0.4587948323856965 accuracy 0.7931711979609176
Val   loss 0.3846700458902963 accuracy 0.8738451736221726

Epoch 2/10
----------

100%|█████████████████████████████████████| 14124/14124 [05:21<00:00, 43.93it/s]

Train loss 0.3450994613828836 accuracy 0.8979219767771169
Val   loss 0.39124348195663816 accuracy 0.8932781140490602

Epoch 3/10
----------

100%|█████████████████████████████████████| 14124/14124 [05:14<00:00, 44.93it/s]

Train loss 0.3135276534462201 accuracy 0.9156046445766072
Val   loss 0.38953639226077036 accuracy 0.9041095890410958

Epoch 4/10
----------

100%|█████████████████████████████████████| 14124/14124 [04:32<00:00, 51.76it/s]

Train loss 0.29076329547278607 accuracy 0.926224865477202
Val   loss 0.4083191853780146 accuracy 0.9063395985982797

Epoch 5/10
----------

100%|█████████████████████████████████████| 14124/14124 [04:33<00:00, 51.70it/s]

Train loss 0.2712314691068196 accuracy 0.9351989521382045
Val   loss 0.44957431750859633 accuracy 0.9063395985982797

Epoch 6/10
----------

100%|█████████████████████████████████████| 14124/14124 [04:28<00:00, 52.56it/s]

Train loss 0.2521194787317903 accuracy 0.9424561030869442
Val   loss 0.4837963371119771 accuracy 0.9082510353615801

Epoch 7/10
----------

100%|█████████████████████████████████████| 14124/14124 [04:28<00:00, 52.64it/s]

Train loss 0.2317749120263705 accuracy 0.9494831492495044
Val   loss 0.5409662437294889 accuracy 0.9063395985982797

Epoch 8/10
----------

100%|█████████████████████████████████████| 14124/14124 [04:29<00:00, 52.39it/s]

Train loss 0.2093608888886245 accuracy 0.9562269895213821
Val   loss 0.5704389385299592 accuracy 0.9037910162472125

Epoch 9/10
----------

100%|█████████████████████████████████████| 14124/14124 [04:28<00:00, 52.68it/s]

Train loss 0.1867563983566425 accuracy 0.9619088077032002
Val   loss 0.6150021497048127 accuracy 0.9015610066900287

Epoch 10/10
----------

100%|█████████████████████████████████████| 14124/14124 [04:29<00:00, 52.45it/s]

Train loss 0.16439846786478746 accuracy 0.9669003115264797
Val   loss 0.6261858006026605 accuracy 0.9098438993309972

使用Conv2D 的效果比Conv1D的效果好。

最近在忙着打一个数据挖掘的比赛，后续会持续输出，请大家关注，谢谢！