Class TextRCNN

model/classification/textrcnn.py:23–91

TextRNN + TextCNN

class TextRCNN(Classifier):
    """TextRNN + TextCNN
    """
    def __init__(self, dataset, config):
        super(TextRCNN, self).__init__(dataset, config)
        self.rnn = RNN(
            config.embedding.dimension, config.TextRCNN.hidden_dimension,
            num_layers=config.TextRCNN.num_layers,
            batch_first=True, bidirectional=config.TextRCNN.bidirectional,
            rnn_type=config.TextRCNN.rnn_type)

        hidden_dimension = config.TextRCNN.hidden_dimension
        if config.TextRCNN.bidirectional:
            hidden_dimension *= 2
        self.kernel_sizes = config.TextRCNN.kernel_sizes
        self.convs = torch.nn.ModuleList()
        for kernel_size in self.kernel_sizes:
            self.convs.append(torch.nn.Conv1d(
                hidden_dimension, config.TextRCNN.num_kernels,
                kernel_size, padding=kernel_size - 1))

        self.top_k = self.config.TextRCNN.top_k_max_pooling
        hidden_size = len(config.TextRCNN.kernel_sizes) * \
                      config.TextRCNN.num_kernels * self.top_k

        self.linear = torch.nn.Linear(hidden_size, len(dataset.label_map))
        self.dropout = torch.nn.Dropout(p=config.train.hidden_layer_dropout)

    def get_parameter_optimizer_dict(self):
        params = list()
        params.append({'params': self.token_embedding.parameters()})
        params.append({'params': self.char_embedding.parameters()})
        params.append({'params': self.rnn.parameters()})
        params.append({'params': self.convs.parameters()})
        params.append({'params': self.linear.parameters()})
        return params

    def update_lr(self, optimizer, epoch):
        """
        """
        if epoch > self.config.train.num_epochs_static_embedding:
            for param_group in optimizer.param_groups[:2]:
                param_group["lr"] = self.config.optimizer.learning_rate
        else:
            for param_group in optimizer.param_groups[:2]:
                param_group["lr"] = 0

    def forward(self, batch):
        if self.config.feature.feature_names[0] == "token":
            embedding = self.token_embedding(
                batch[cDataset.DOC_TOKEN].to(self.config.device))
            seq_length = batch[cDataset.DOC_TOKEN_LEN].to(self.config.device)
        else:
            embedding = self.char_embedding(
                batch[cDataset.DOC_CHAR].to(self.config.device))
            seq_length = batch[cDataset.DOC_CHAR_LEN].to(self.config.device)
        output, _ = self.rnn(embedding, seq_length)