Aug 18, 2017 by Li Jing

In this post, I will illustrate recurrent neural network (RNN) Encoder-Decoder with the input and output. The model architecture is based on the official tutorial of Pytorch.

Illustration of input and output:

Figure: Illustration input and output of RNN Encoder-Decoder.
bmm: The context vector `c_i = \sum_{j=1}^{T_x} \alpha_{ij}h_j` in paper Bahdanau et al.
In above figure, for time step `t` of decoder: $$c_t = [\alpha_{t1}, \alpha_{t2},..., \alpha_{tL}]_{1\times L} \times \left[ \begin{array}{l} h_1\\ h_2\\ h_3\\ h_L \end{array} \right]_{L \times H}$$

Code


import torch
import torch.nn as nn
from torch.autograd import Variable
from torch import optim
import torch.nn.functional as F
from seq2seq.data_helper import use_cuda,MAX_LENGTH


class EncoderRNN(nn.Module):
    def __init__(self, input_size, hidden_size, embed_size, n_layers=1):
        super(EncoderRNN, self).__init__()
        self.n_layers = n_layers
        self.hidden_size = hidden_size

        self.embedding = nn.Embedding(input_size, embed_size)
        self.gru = nn.GRU( embed_size,hidden_size)

    def forward(self, input, hidden):
        embedded = self.embedding(input).view(1, 1, -1)
        output = embedded
        for i in range(self.n_layers):
            output, hidden = self.gru(output, hidden)
        return output, hidden

    def initHidden(self):
        result = Variable(torch.zeros(1, 1, self.hidden_size))
        if use_cuda:
            return result.cuda()
        else:
            return result




class AttnDecoderRNN(nn.Module):
    def __init__(self, hidden_size, output_size, embed_size, n_layers=1, dropout_p=0.1, max_length=MAX_LENGTH):
        super(AttnDecoderRNN, self).__init__()
        self.hidden_size = hidden_size
        self.output_size = output_size
        self.n_layers = n_layers
        self.dropout_p = dropout_p
        self.max_length = max_length
        self.embed_size = embed_size

        self.embedding = nn.Embedding(self.output_size, self.embed_size)
        self.attn = nn.Linear(self.hidden_size  + self.embed_size, self.max_length)
        self.attn_combine = nn.Linear(self.hidden_size + self.embed_size, self.hidden_size)
        self.dropout = nn.Dropout(self.dropout_p)
        self.gru = nn.GRU(self.hidden_size, self.hidden_size)
        self.out = nn.Linear(self.hidden_size, self.output_size)

    def forward(self, input, hidden, encoder_output, encoder_outputs):
        embedded = self.embedding(input).view(1, 1, -1)
        embedded = self.dropout(embedded)

        attn_weights = F.softmax(
            self.attn(torch.cat((embedded[0], hidden[0]), 1)))
        attn_applied = torch.bmm(attn_weights.unsqueeze(0),
                                 encoder_outputs.unsqueeze(0))

        output = torch.cat((embedded[0], attn_applied[0]), 1)
        output = self.attn_combine(output).unsqueeze(0)

        for i in range(self.n_layers):
            output = F.relu(output)
            output, hidden = self.gru(output, hidden)

        output = F.log_softmax(self.out(output[0]))
        return output, hidden, attn_weights

    def initHidden(self):
        result = Variable(torch.zeros(1, 1, self.hidden_size))
        if use_cuda:
            return result.cuda()
        else:
            return result