sundea/text-classification

1# coding: UTF-82import torch3import torch.nn as nn4import torch.nn.functional as F5import numpy as np6 7 8class Config(object):9 10    """配置参数"""11    def __init__(self, dataset, embedding):12        self.model_name = 'DPCNN'13        self.train_path = dataset + '/data/train.txt'                                # 训练集14        self.dev_path = dataset + '/data/dev.txt'                                    # 验证集15        self.test_path = dataset + '/data/test.txt'                                  # 测试集16        self.class_list = [x.strip() for x in open(17            dataset + '/data/class.txt', encoding='utf-8').readlines()]              # 类别名单18        self.vocab_path = dataset + '/data/vocab.pkl'                                # 词表19        self.save_path = dataset + '/saved_dict/' + self.model_name + '.ckpt'        # 模型训练结果20        self.log_path = dataset + '/log/' + self.model_name21        self.embedding_pretrained = torch.tensor(22            np.load(dataset + '/data/' + embedding)["embeddings"].astype('float32'))\23            if embedding != 'random' else None                                       # 预训练词向量24        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')   # 设备25 26        self.dropout = 0.5                                              # 随机失活27        self.require_improvement = 1000                                 # 若超过1000batch效果还没提升，则提前结束训练28        self.num_classes = len(self.class_list)                         # 类别数29        self.n_vocab = 0                                                # 词表大小，在运行时赋值30        self.num_epochs = 20                                            # epoch数31        self.batch_size = 128                                           # mini-batch大小32        self.pad_size = 32                                              # 每句话处理成的长度(短填长切)33        self.learning_rate = 1e-3                                       # 学习率34        self.embed = self.embedding_pretrained.size(1)\35            if self.embedding_pretrained is not None else 300           # 字向量维度36        self.num_filters = 250                                          # 卷积核数量(channels数)37 38 39'''Deep Pyramid Convolutional Neural Networks for Text Categorization'''40 41 42class Model(nn.Module):43    def __init__(self, config):44        super(Model, self).__init__()45        if config.embedding_pretrained is not None:46            self.embedding = nn.Embedding.from_pretrained(config.embedding_pretrained, freeze=False)47        else:48            self.embedding = nn.Embedding(config.n_vocab, config.embed, padding_idx=config.n_vocab - 1)49        self.conv_region = nn.Conv2d(1, config.num_filters, (3, config.embed), stride=1)50        self.conv = nn.Conv2d(config.num_filters, config.num_filters, (3, 1), stride=1)51        self.max_pool = nn.MaxPool2d(kernel_size=(3, 1), stride=2)52        self.padding1 = nn.ZeroPad2d((0, 0, 1, 1))  # top bottom53        self.padding2 = nn.ZeroPad2d((0, 0, 0, 1))  # bottom54        self.relu = nn.ReLU()55        self.fc = nn.Linear(config.num_filters, config.num_classes)56 57    def forward(self, x):58        x = x[0]59        x = self.embedding(x)60        x = x.unsqueeze(1)  # [batch_size, 250, seq_len, 1]61        x = self.conv_region(x)  # [batch_size, 250, seq_len-3+1, 1]62 63        x = self.padding1(x)  # [batch_size, 250, seq_len, 1]64        x = self.relu(x)65        x = self.conv(x)  # [batch_size, 250, seq_len-3+1, 1]66        x = self.padding1(x)  # [batch_size, 250, seq_len, 1]67        x = self.relu(x)68        x = self.conv(x)  # [batch_size, 250, seq_len-3+1, 1]69        while x.size()[2] > 2:70            x = self._block(x)71        x = x.squeeze()  # [batch_size, num_filters(250)]72        x = self.fc(x)73        return x74 75    def _block(self, x):76        x = self.padding2(x)77        px = self.max_pool(x)78 79        x = self.padding1(px)80        x = F.relu(x)81        x = self.conv(x)82 83        x = self.padding1(x)84        x = F.relu(x)85        x = self.conv(x)86 87        # Short Cut88        x = x + px89        return x90