项目1：pytorch实现文本情感分析详细教程-准确度高达82%-98%

比赛链接

数据集介绍

数据集为英文文本数据，其中Label为其情感标签，正负类样本各有12500个。总计样本数量为两万五千条。

数据预处理

获取停用词表，用于过滤不想关的单词

数据分词、词干提取、词性还原、过滤停用词等操作

lemmatizer = nltk.WordNetLemmatizer()
strem = PorterStemmer()
def is_contain_number(word):
    for i in range(10):
        if str(i) in word:
            return True
    return False
def preprocess_sentence(sentence):
    # print("开始分词...")
    processed_tokens = nltk.word_tokenize(sentence)
    # 词干提取
    processed_tokens = [strem.stem(w.lower()) for w in processed_tokens]
    # print("去除部分tokens...")
    processed_tokens2 = [w for w in processed_tokens if ((w not in stop_list) and (not is_contain_number(w)))]
    # print("词性还原...")
    processed_tokens = [lemmatizer.lemmatize(w) for w in processed_tokens2]
    return processed_tokens 
def tokenize(string):
    if not type(string) is str:
        return []
    return preprocess_sentence(string)

加载Glove2Word2vec模型

这里的Glove预训练模型官网下载较慢，私聊我可获取完整资料！

from gensim.scripts.glove2word2vec import glove2word2vec
glove_input_file = 'glove.6B.100d.txt'
word2vec_output_file = 'glove.6B.100d.word2vec.txt'
(count, dimensions) = glove2word2vec(glove_input_file, word2vec_output_file)
print(count, '\n', dimensions)

根据预训练模型的词典对数据进行过滤，并保证数据长度一致，过长的数据将其截断，过短的数据则以“PAD”为填充符。

max_seq_len = 0
for i,sentence in enumerate(tqdm(sentence_list)):
    sent_list = sentence.split()
    sent_list = [x for x in sent_list if x in word2idx.keys()]
    max_seq_len = max(max_seq_len,len(sent_list))
    sentence_list[i] = " ".join(sent_list)
print(max_seq_len)
max_seq_len = min(max_seq_len,200)
print(max_seq_len)

如下所示：

将单词转化为字典序号

import torch 
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
def make_data(sentences, labels):
    inputs = []
    for sen in tqdm(sentences):
        inputs.append([word2idx[n] for n in sen.split()])

    targets = labels.tolist()
    return np.array(inputs), np.array(targets)

input_batch, target_batch = make_data(sentence_list, labels)

构建模型

class Bi_Lstm(nn.Module):
    def __init__(self):
        super(Bi_Lstm,self).__init__() 
        self.embeddings = torch.tensor(glove_model.vectors)
        self.add_embedding('')
        self.embeding = nn.Embedding(self.embeddings.shape[0],self.embeddings.shape[1], padding_idx=word2idx[''])
        self.embeding.weight.data.copy_(self.embeddings)
        self.embeding.weight.requires_grad = False
        self.lstm = nn.LSTM(input_size = self.embeddings.shape[1], hidden_size = 100,num_layers = 2, bidirectional = True, batch_first=True, dropout=0.3)#加了双向，输出的节点数翻2倍
        self.l1 = nn.BatchNorm1d(self.embeddings.shape[1]*2)
        self.l2 = nn.ReLU()
        self.dense = nn.Sequential(
            nn.Linear(self.embeddings.shape[1]*2,128),
            nn.BatchNorm1d(128),
            nn.Dropout(0.2),
            nn.Linear(128,64),
            nn.BatchNorm1d(64),
            nn.Linear(64,n_class),
            nn.BatchNorm1d(n_class)
        )

    def add_embedding(self, word):
        vector = torch.empty(1, self.embeddings.shape[1])#生成空的
        torch.nn.init.uniform_(vector)#随机生成
        self.embeddings = torch.cat([self.embeddings, vector], 0)#在embedding_matrix中加入新的vector

    def forward(self, x):
        x = self.embeding(x)
        out,_ = self.lstm(x)
        #选择最后一个时间点的output
        out = self.l1(out[:,-1,:])
        out = self.l2(out)
        out = self.dense(out)
        return out

开始训练

from sklearn.metrics import classification_report, confusion_matrix
import warnings
warnings.filterwarnings("ignore")
training_step = 50#迭代次数
batch_size = 512#每个批次的大小
kf = KFold(n_splits=5)#5折交叉验证
def res2array(xx):
    data = []
    [data.extend(x.tolist()) for x in xx]
    return np.array(data)
for fold, (train_idx, test_idx) in enumerate(kf.split(train, targets)):
    max_acc = 0.0
    print('-'*15, '>', f'Fold {fold+1}', '<', '-'*15)
    x_train, x_val = train[train_idx], train[test_idx]
    y_train, y_val = targets[train_idx], targets[test_idx]
    
    M_train = len(x_train)-1

    x_train = torch.from_numpy(x_train).to(torch.long).to(device)
    x_val = torch.from_numpy(x_val).to(torch.long).to(device)
    y_train = torch.from_numpy(y_train).to(torch.long).to(device)
    y_val = torch.from_numpy(y_val).to(torch.long).to(device)

    model = Bi_Lstm()
    model.to(device)
    optimizer = torch.optim.Adam(model.parameters(),lr=0.001)
    loss_func = nn.CrossEntropyLoss()#多分类的任务
    model.train()#模型中有BN和Droupout一定要添加这个说明
    #开始迭代
    for step in range(training_step):
        ytrain_pre_epoch = []
        ytrain_true_epoch = []
        yval_pre_epoch = []
        yval_true_epoch = []
        train_acc_per_epoch = 0.0
        val_acc_per_epoch = 0.0
        train_loss_per_epoch = 0.0
        val_loss_per_epoch = 0.0
        all_step = 0
        with tqdm(np.arange(0,M_train,batch_size), desc='Training...') as tbar:
            for index in tbar:
                L = index
                R = min(M_train,index+batch_size)
                if (R-L) <= 1:
                    continue

                #-----------------训练内容------------------
                model.train()
                optimizer.zero_grad()   # 清空上一步的残余更新参数值
                train_pre = model(x_train[L:R])     # 喂给 model训练数据 x, 输出预测值
                train_loss = loss_func(train_pre, y_train[L:R])

                
                #-----------------反向传播更新---------------
                train_loss.backward()         # 以训练集的误差进行反向传播, 计算参数更新值
                optimizer.step()        # 将参数更新值施加到 net 的 parameters 上

                #----------- -----计算准确率----------------
                ytrain_pre = np.argmax(np.array(train_pre.data.cpu()),axis=1)
                train_true = np.array(y_train[L:R].data.cpu())
                ytrain_pre_epoch.append(ytrain_pre)
                ytrain_true_epoch.append(train_true)
                train_acc = np.sum(np.argmax(np.array(train_pre.data.cpu()),axis=1) == np.array(y_train[L:R].data.cpu()))/(R-L)

                #---------------打印在进度条上--------------
                tbar.set_postfix(
                    train_loss=float(train_loss.data.cpu()),
                    train_acc=train_acc
                    )
                tbar.update()  # 默认参数n=1，每update一次，进度+n
                train_acc_per_epoch += train_acc
                train_loss_per_epoch += train_loss.data.cpu()
                all_step += 1

        # --------------------每个epoch结束就要验证一次模型的准确度-----------------------------
        with torch.no_grad():
            M_val = len(x_val) - 1
            all_step = 0
            with tqdm(np.arange(0,M_val,batch_size), desc='Validation...') as tbar:
                for index in tbar:
                    L_val = index
                    R_val = min(M_val,L_val + batch_size)
                    if (R_val-L_val) <= 1:
                        continue
                    
                    #-----------------验证------------------
                    model.eval()
                    val_pre = model(x_val[L_val:R_val])#验证集也得分批次，不然数据量太大内存爆炸
                    val_loss = loss_func(val_pre, y_val[L_val:R_val])

                    yval_pre = np.argmax(np.array(val_pre.data.cpu()),axis=1)
                    val_true = np.array(y_val[L_val:R_val].data.cpu())
                    
                    yval_pre_epoch.append(yval_pre)
                    yval_true_epoch.append(val_true)
                    
                    val_acc = np.sum(np.argmax(np.array(val_pre.data.cpu()),axis=1) == np.array(y_val[L_val:R_val].data.cpu()))/(R_val-L_val)
                    
                    #---------------打印在进度条上--------------
                    tbar.set_postfix(
                        val_loss=float(val_loss.data.cpu()),
                        val_acc=val_acc
                        )
                    tbar.update()  # 默认参数n=1，每update一次，进度+n
                    val_acc_per_epoch += val_acc
                    val_loss_per_epoch += val_loss.data.cpu()
                    all_step += 1

        res = classification_report(res2array(ytrain_pre_epoch),res2array(ytrain_true_epoch),digits=3,output_dict=True)
        # print(res)
        train_stat = res["weighted avg"] 
        train_stat["accuracy"] = res["accuracy"]
        res = classification_report(res2array(yval_pre_epoch),res2array(yval_true_epoch),digits=3,output_dict=True)
        # print(res)
        val_stat = res["weighted avg"] 
        val_stat["accuracy"] = res["accuracy"]
        print("Epoch %d/%d: trainning acc: %.3f validation acc: %.3f"%(step,training_step,train_stat["accuracy"],val_stat["accuracy"]))
        model_path = "model_fold_"+str(fold)+".pt"
        if res["accuracy"] > max_acc:
            max_acc = res["accuracy"]
            torch.save(model,model_path)

提交测试结果