modeling_bilstm.py

import torch
import torch.nn as nn
from transformers import PretrainedConfig, PreTrainedModel

class BiLSTMSequenceLabelerConfig(PretrainedConfig):
    model_type = "bilstm_sequence_labeler"
    def __init__(self, vocab_size=30000, embedding_dim=300, hidden_dim=128, num_ner_classes=9, num_pos_classes=45, num_chunk_classes=23, **kwargs):
        super().__init__(**kwargs)
        self.vocab_size = vocab_size
        self.embedding_dim = embedding_dim
        self.hidden_dim = hidden_dim
        self.num_ner_classes = num_ner_classes
        self.num_pos_classes = num_pos_classes
        self.num_chunk_classes = num_chunk_classes

class HFBiLSTMSequenceLabeler(PreTrainedModel):
    config_class = BiLSTMSequenceLabelerConfig
    def __init__(self, config):
        super().__init__(config)
        self.embedding = nn.Embedding(config.vocab_size, config.embedding_dim, padding_idx=0)
        self.rnn = nn.LSTM(config.embedding_dim, config.hidden_dim, batch_first=True, bidirectional=True)
        rnn_out_dim = config.hidden_dim * 2
        self.ner_head = nn.Linear(rnn_out_dim, config.num_ner_classes)
        self.pos_head = nn.Linear(rnn_out_dim, config.num_pos_classes)
        self.chunk_head = nn.Linear(rnn_out_dim, config.num_chunk_classes)
        # CRUCIAL: Hook for newer transformers compatibility (initializes all_tied_weights_keys)
        self.post_init()

    def forward(self, input_ids, **kwargs):
        embedded = self.embedding(input_ids)
        rnn_out, _ = self.rnn(embedded)
        return {
            "ner": self.ner_head(rnn_out),
            "pos": self.pos_head(rnn_out),
            "chunk": self.chunk_head(rnn_out)
        }