TextCNN的复现–pytorch的实现
对于TextCNN的讲解，可以参考这篇文章
Convolutional Neural Networks for Sentence Classification – 知乎 (zhihu.com)
接下来主要是对代码内容的详解，完整代码将在文章末尾给出。
使用的数据集为电影评论数据集，其中正面数据集5000条左右，负面的数据集也为5000条。
pyroch的基本训练过程：
加载训练集–构建模型–模型训练–模型评价
首先，是要对数据集进行加载，在对数据集加载时候需要继承一下Dataset类，代码如下
class Data_loader(Dataset):def __init__(self, file_pos, file_neg, model_path, word2_vec=False):self.file_pos = file_posself.file_neg = file_negif word2_vec:self.x_train, self.y_train = self.get_word2vec(model_path)else:self.x_train, self.y_train, self.dictionary = self.pre_process()def __getitem__(self, idx):data = self.x_train[idx]label = self.y_train[idx]data = torch.tensor(data)label = torch.tensor(label)return data, labeldef __len__(self):return len(self.x_train)def clean_sentences(self, string):string = re.sub(r"[^A-Za-z0-9(),!?\'\`]", " ", string)string = re.sub(r"\'s", " \'s", string)string = re.sub(r"\'ve", " \'ve", string)string = re.sub(r"n\'t", " n\'t", string)string = re.sub(r"\'re", " \'re", string)string = re.sub(r"\'d", " \'d", string)string = re.sub(r"\'ll", " \'ll", string)string = re.sub(r",", " , ", string)string = re.sub(r"!", " ! ", string)string = re.sub(r"\(", " \( ", string)string = re.sub(r"\)", " \) ", string)string = re.sub(r"\?", " \? ", string)string = re.sub(r"\s{2,}", " ", string)return string.strip().lower()def load_data_and_labels(self):positive_examples = list(open(self.file_pos, "r", encoding="utf-8").readlines())positive_examples = [s.strip() for s in positive_examples]# 对评论数据删除每一行数据的\t，\nnegative_examples = list(open(self.file_neg, "r", encoding="utf-8").readlines())negative_examples = [s.strip() for s in negative_examples]# 对评论数据删除每一行数据的\t，\nx_text = positive_examples + negative_examplesx_text = [self.clean_sentences(_) for _ in x_text]positive_labels = [[1, 0] for _ in positive_examples]# 正样本数据为1negative_labels = [[0, 1] for _ in negative_examples]# 负样本数据为0y = np.concatenate([positive_labels, negative_labels], 0)return x_text, y# 返回的是dataframe对象，[0]data[0]为文本数据，data[1]为标签def pre_process(self):'''加载数据，并对之前使用的数据进行打乱返回,同时根据训练集和测试集的比列进行划分,默认百分80和百分20:return:测试数据、训练数据、以及生成的词汇表'''x_data, y_label = self.load_data_and_labels()max_document_length = max(len(x.split(' ')) for x in x_data)voc = []word_split = [][voc.extend(x.split()) for x in x_data]# 生成词典[word_split.append(x.split()) for x in x_data]if len(voc) != 0:ordere_dict = OrderedDict(sorted(Counter(_flatten(voc)).items(), key=lambda x: x[1], reverse=True))# 把文档映射成词汇的索引序列dictionary = vocab(ordere_dict)x_data = []for words in word_split:x = list(dictionary.lookup_indices(words))temp_pos = max_document_length - len(x)if temp_pos != 0:for i in range(1, temp_pos + 1):x.extend([0])x_data.append(x)x_data = np.array(x_data)np.random.seed(10)# 将标签打乱顺序，返回索引shuffle_indices = np.random.permutation(np.arange(len(y_label)))x_shuffled = x_data[shuffle_indices]y_shuffled = y_label[shuffle_indices]return x_shuffled, y_shuffled, dictionarydef get_word2vec(self, model_path):model = gensim.models.Word2Vec.load(model_path)x_data, y_label = self.load_data_and_labels()word_split = [][word_split.append(x.split()) for x in x_data]sentence_vectors = []for sentence in word_split:sentence_vector = []for word in sentence:try:v = model.wv.get_vector(word)except Exception as e:v = np.zeros(shape=(model.vector_size,), dtype=np.float32)sentence_vector.append(v)sentence_vectors.append(sentence_vector)max_document_length = max(len(x) for x in sentence_vectors)for vector in sentence_vectors:for i in range(1, max_document_length - len(vector) + 1):v = np.zeros(shape=(model.vector_size,), dtype=np.float32)vector.append(v)vector_data = np.asarray(sentence_vectors, dtype=np.float32)np.random.seed(10)# 将标签打乱顺序，返回索引shuffle_indices = np.random.permutation(np.arange(len(y_label)))x_shuffled = vector_data[shuffle_indices]y_shuffled = y_label[shuffle_indices]return x_shuffled, y_shuffled上述代码中的__init__ 、getitem 、len是必须要继承实现的方法，clean_sentence是对读取的数据进行清洗，load_data_and_label是加载数据且返回清洗过后的数据以及数据标签。pre_process是对数据进行编码，原始的数据是英文数据，因此需要对其进行分词、编码，最后返回的数据将是数字，一行数据就是一句评论。
例如：
I like this movie
在对其进行编码返回后将是 0 1 2 3,0对应的为I，1对应的为like以此类推。
get_word2vec则是使用word2vec预训练模型来对每个单词对应的数据内容进行映射。1个单词对应的将会是一个100维的矩阵，该维度可以根据自己训练word2vec模型时候自己进行调整。
接下来是word2vec模型的训练及保存，出于简便性，训练word2vec模型时候直接使用了该数据集对word2vec模型进行训练。
代码如下所示:
def get_model(p_file, n_file):x_data, y_label = load_data_and_labels(p_file, n_file)x_data = [x.split() for x in x_data]max_document_length = max(len(x) for x in x_data)model = Word2Vec(x_data, vector_size=256)return model在这儿设置的每个词的维度是256维。
接下来就是TextCNN模型的构建
class GlobalMaxPool1d(nn.Module):def __init__(self):super(GlobalMaxPool1d, self).__init__()def forward(self, x):return F.max_pool1d(x, kernel_size=x.shape[2])class TextCNN(nn.Module):def __init__(self, num_classes, num_embeddings=-1, embedding_dim=512, kernel_size=[3, 4, 5, 6], num_channels=[32, 32, 32, 32], embeddings_pretrained=None):super(TextCNN, self).__init__()self.num_classes = num_classesself.num_embeddings = num_embeddingsif self.num_embeddings > 0:self.embedding = nn.Embedding(num_embeddings, embedding_dim)if embeddings_pretrained is not None:self.embedding = self.embedding.from_pretrained(embeddings_pretrained, freeze=False)self.cnn_layers = nn.ModuleList()# 创建多个一维卷积层for c, k in zip(num_channels, kernel_size):cnn = nn.Sequential(nn.Conv1d(in_channels=embedding_dim, out_channels=c, kernel_size=k),nn.BatchNorm1d(c),nn.ReLU(inplace=True))# cnn = nn.Sequential(nn.Conv2d(in_channels=1, out_channels=c, kernel_size=k),# nn.BatchNorm1d(c),# nn.ReLU(inplace=True)# )self.cnn_layers.append(cnn)self.pool = GlobalMaxPool1d()self.classify = nn.Sequential(nn.Dropout(p=0.2),nn.Linear(sum(num_channels), self.num_classes))def forward(self, x):if self.num_embeddings > 0:x = self.embedding(x)# input = torch.unsqueeze(x, dim=1)# print(input.size())input = x.permute(0, 2, 1)# print(input.size())# print(len(input[0]))y = []for layer in self.cnn_layers:x = layer(input)x = self.pool(x).squeeze(-1)y.append(x)# print(y)y = torch.cat(y, dim=1)out = self.classify(y)# out = torch.sigmoid(out)return out在构建模型时候需要继承nn.moudule,同时要实现__init__、以及forward方法，可以看作init在定义各个层，forward在对各个层之间来进行连接。
接下来就是对模型进行训练，代码如下所示:
batch_size = 832num_classes = 2file_pos = 'E:\\PostGraduate\\Paper_review\\pytorch_TextCnn/data/rt-polarity.pos'file_neg = 'E:\\PostGraduate\\Paper_review\\pytorch_TextCnn/data/rt-polarity.neg'word2vec_path = 'E:\\PostGraduate\\Paper_review\\pytorch_TextCnn/word2vec1.model'train_data = Data_loader(file_pos, file_neg, word2vec_path)train_size = int(len(train_data) * 0.8)test_size = len(train_data) - train_sizetrain_dataset, test_dataset = torch.utils.data.random_split(train_data, [train_size, test_size])train_iter = DataLoader(train_dataset, batch_size=830, shuffle=True)test_iter = DataLoader(test_dataset, batch_size=2133, shuffle=True)model = TextCNN(num_classes, embeddings_pretrained=True)# model = TextCNN(num_classes, num_embeddings=18764)# 开始训练epoch = 100# 训练轮次optmizer = torch.optim.Adam(model.parameters(), lr=0.01)# optmizer = torch.optim.SGD(model.parameters(),lr=0.01,momentum=0.4)train_losses = []train_counter = []test_losses = []log_interval = 5test_counter = [i * len(train_iter.dataset) for i in range(epoch + 1)]device = 'cpu'def train_loop(n_epochs, optimizer, model, train_loader, device, test_iter):for epoch in range(1, n_epochs + 1):print("开始第{}轮训练".format(epoch))model.train()correct = 0for i, data in enumerate(train_loader):optimizer.zero_grad()(text_data, label) = datatext_data = text_data.to(device)label = label.to(device)label = label.long()output = model(text_data)loss_func = nn.BCEWithLogitsLoss()# output = output.long()loss = loss_func(output, label.float())loss.backward()optimizer.step()pred = output.data.max(1, keepdim=True)[1]label = label.data.max(1, keepdim=True)[1]correct += pred.eq(label.data.view_as(pred)).sum()if i % log_interval == 0:print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(epoch, i * len(text_data), len(train_loader.dataset), 100. * i / len(train_loader), loss.item()))train_losses.append(loss.item())train_counter.append((i * 64) + ((epoch - 1) * len(train_loader.dataset)))torch.save(model.state_dict(), './model.pth')torch.save(optimizer.state_dict(), './optimizer.pth')print("Accuracy: {}/{} ({:.0f}%)\n".format(correct, len(train_loader.dataset), 100. * correct / len(train_loader.dataset)))print("开始第{}轮评价".format(epoch))model.eval()test_loss = 0correct = 0with torch.no_grad():for data, target in test_iter:# for data, target in train_iter:data = data.to(device)target = target.to(device)output = model(data)loss_func = nn.BCEWithLogitsLoss()# output = output.long()loss = loss_func(output, target.float())test_loss += losspred = output.data.max(1, keepdim=True)[1]label = target.data.max(1, keepdim=True)[1]correct += pred.eq(label.data.view_as(pred)).sum()test_loss /= len(test_iter.dataset)# test_loss /= len(train_iter.dataset)test_losses.append(test_loss)print('\nTest set: Avg. loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(test_loss, correct, len(test_iter.dataset),100. * correct / len(test_iter.dataset)))train_loop(epoch, optmizer, model, train_iter, device, test_iter)在上述中，首先会对数据集加载进来，然后分为80%的训练集和20%的测试集，定义使用的优化器为adam。同时在训练的过程中会对优化器、损失函数等信息进行保存。
训练结果如下所示：
完整代码链接
t, len(test_iter.dataset),
100. * correct / len(test_iter.dataset)))
train_loop(epoch, optmizer, model, train_iter, device, test_iter)
在上述中，首先会对数据集加载进来，然后分为80%的训练集和20%的测试集，定义使用的优化器为adam。同时在训练的过程中会对优化器、损失函数等信息进行保存。训练结果为75%左右。完整代码链接[木南/TextCNN (gitee.com)](https://gitee.com/nanwang-crea/text-cnn)