darshan204/HumanValuesUncover

1import torch2import torch.nn as nn3 4class BiLSTMClassifier(nn.Module):5    def __init__(self, embedding_dim, hidden_dim, output_dim, n_layers, bidirectional, dropout):6        super().__init__()7        self.embedding_dim = embedding_dim8        self.hidden_dim = hidden_dim9        self.output_dim = output_dim10        self.n_layers = n_layers11        self.bidirectional = bidirectional12        self.dropout = dropout13 14        self.lstm = nn.LSTM(embedding_dim,15                            hidden_dim,16                            num_layers=n_layers,17                            bidirectional=bidirectional,18                            dropout=dropout,19                            batch_first=True)20 21        self.fc = nn.Linear(hidden_dim * 2 if bidirectional else hidden_dim, output_dim)22        self.dropout = nn.Dropout(dropout)23 24    def forward(self, text, text_lengths):25        # text = [batch size, sent len]26 27        # pack sequence28        packed_embedded = nn.utils.rnn.pack_padded_sequence(text, text_lengths.cpu(), batch_first=True, enforce_sorted=False)29 30        packed_output, (hidden, cell) = self.lstm(packed_embedded)31 32        # unpack sequence33        output, output_lengths = nn.utils.rnn.pad_packed_sequence(packed_output, batch_first=True)34 35        # output = [batch size, sent len, hidden dim * n directions]36        # hidden = [n layers * n directions, batch size, hidden dim]37 38        # concat the final forward (hidden[-2,:,:]) and backward (hidden[-1,:,:]) hidden layers39        # and apply dropout40        hidden = self.dropout(torch.cat((hidden[-2,:,:], hidden[-1,:,:]), dim=1) if self.bidirectional else hidden[-1,:,:])41 42        # hidden = [batch size, hidden dim * n directions]43 44        return self.fc(hidden)