Example of LSTM encoder which ignores labels for padding

LSTMLEncoder.py

import chainer
from chainer import Link, Chain, Function, Variable, cuda
import chainer.functions as F
import chainer.links as L
import numpy as np

class LSTMEncoder(Chain):
    def __init__(self, vocab_size, embed_size, hidden_size, ignore_label=-1):
        super(LSTMEncoder, self).__init__(
            xe = L.EmbedID(vocab_size, embed_size, ignore_label=-1),
            eh = L.Linear(embed_size, 4 * hidden_size),
            hh = L.Linear(hidden_size, 4 * hidden_size),
        )
        self.ignore_label = ignore_label

    def __call__(self, x, c_prev, h_prev):
        xp = cuda.get_array_module(x.data)
        e = F.tanh(self.xe(x))
        c, h = F.lstm(c_prev, self.eh(e) + self.hh(h_prev))
        ignore = xp.broadcast_to(xp.reshape((x.data == self.ignore_label), (x.shape[0], 1)), c.shape)
        c_next = F.where(ignore, c_prev, c)
        h_next = F.where(ignore, h_prev, h)
        return c_next, h_next