mmocr/models/textrecog/layers/lstm_layer.py

# Copyright (c) OpenMMLab. All rights reserved.
import torch.nn as nn


class BidirectionalLSTM(nn.Module):

    def __init__(self, nIn, nHidden, nOut):
        super().__init__()

        self.rnn = nn.LSTM(nIn, nHidden, bidirectional=True)
        self.embedding = nn.Linear(nHidden * 2, nOut)

    def forward(self, input):
        recurrent, _ = self.rnn(input)
        T, b, h = recurrent.size()
        t_rec = recurrent.view(T * b, h)

        output = self.embedding(t_rec)  # [T * b, nOut]
        output = output.view(T, b, -1)

        return output