# code by Tae Hwan Jung @graykode

import numpy as np

import torch

import torch.nn as nn

  

# S: Symbol that shows starting of decoding input

# E: Symbol that shows starting of decoding output

# P: Symbol that will fill in blank sequence if current batch data size is short than time steps

  

def make_batch():

    input_batch, output_batch, target_batch = [], [], []

  

    for seq in seq_data:

        for i in range(2):

            seq[i] = seq[i] + 'P' * (n_step - len(seq[i]))

  

        input = [num_dic[n] for n in seq[0]]

        output = [num_dic[n] for n in ('S' + seq[1])]

        target = [num_dic[n] for n in (seq[1] + 'E')]

  

        input_batch.append(np.eye(n_class)[input])

        output_batch.append(np.eye(n_class)[output])

        target_batch.append(target) # not one-hot

  

    # make tensor

    return torch.FloatTensor(input_batch), torch.FloatTensor(output_batch), torch.LongTensor(target_batch)

  

# make test batch

def make_testbatch(input_word):

    input_batch, output_batch = [], []

  

    input_w = input_word + 'P' * (n_step - len(input_word))

    input = [num_dic[n] for n in input_w]

    output = [num_dic[n] for n in 'S' + 'P' * n_step]

  

    input_batch = np.eye(n_class)[input]

    output_batch = np.eye(n_class)[output]

  

    return torch.FloatTensor(input_batch).unsqueeze(0), torch.FloatTensor(output_batch).unsqueeze(0)

  

# Model

class Seq2Seq(nn.Module):

    def __init__(self):

        super(Seq2Seq, self).__init__()

  

        self.enc_cell = nn.RNN(input_size=n_class, hidden_size=n_hidden, dropout=0.5)

        self.dec_cell = nn.RNN(input_size=n_class, hidden_size=n_hidden, dropout=0.5)

        self.fc = nn.Linear(n_hidden, n_class)

  

    def forward(self, enc_input, enc_hidden, dec_input):

        enc_input = enc_input.transpose(0, 1) # enc_input: [max_len(=n_step, time step), batch_size, n_class]

        dec_input = dec_input.transpose(0, 1) # dec_input: [max_len(=n_step, time step), batch_size, n_class]

  

        # enc_states : [num_layers(=1) * num_directions(=1), batch_size, n_hidden]

        _, enc_states = self.enc_cell(enc_input, enc_hidden)

        # outputs : [max_len+1(=6), batch_size, num_directions(=1) * n_hidden(=128)]

        outputs, _ = self.dec_cell(dec_input, enc_states)

  

        model = self.fc(outputs) # model : [max_len+1(=6), batch_size, n_class]

        return model

  

if __name__ == '__main__':

    n_step = 5

    n_hidden = 128

  

    char_arr = [c for c in 'SEPabcdefghijklmnopqrstuvwxyz']

    num_dic = {n: i for i, n in enumerate(char_arr)}

    seq_data = [['man', 'women'], ['black', 'white'], ['king', 'queen'], ['girl', 'boy'], ['up', 'down'], ['high', 'low']]

  

    n_class = len(num_dic)

    batch_size = len(seq_data)

  

    model = Seq2Seq()

  

    criterion = nn.CrossEntropyLoss()

    optimizer = torch.optim.Adam(model.parameters(), lr=0.001)

  

    input_batch, output_batch, target_batch = make_batch()

  

    for epoch in range(5000):

        # make hidden shape [num_layers * num_directions, batch_size, n_hidden]

        hidden = torch.zeros(1, batch_size, n_hidden)

  

        optimizer.zero_grad()

        # input_batch : [batch_size, max_len(=n_step, time step), n_class]

        # output_batch : [batch_size, max_len+1(=n_step, time step) (becase of 'S' or 'E'), n_class]

        # target_batch : [batch_size, max_len+1(=n_step, time step)], not one-hot

        output = model(input_batch, hidden, output_batch)

        # output : [max_len+1, batch_size, n_class]

        output = output.transpose(0, 1) # [batch_size, max_len+1(=6), n_class]

        loss = 0

        for i in range(0, len(target_batch)):

            # output[i] : [max_len+1, n_class, target_batch[i] : max_len+1]

            loss += criterion(output[i], target_batch[i])

        if (epoch + 1) % 1000 == 0:

            print('Epoch:', '%04d' % (epoch + 1), 'cost =', '{:.6f}'.format(loss))

        loss.backward()

        optimizer.step()

  

    # Test

    def translate(word):

        input_batch, output_batch = make_testbatch(word)

  

        # make hidden shape [num_layers * num_directions, batch_size, n_hidden]

        hidden = torch.zeros(1, 1, n_hidden)

        output = model(input_batch, hidden, output_batch)

        # output : [max_len+1(=6), batch_size(=1), n_class]

  

        predict = output.data.max(2, keepdim=True)[1] # select n_class dimension

        decoded = [char_arr[i] for i in predict]

        end = decoded.index('E')

        translated = ''.join(decoded[:end])

  

        return translated.replace('P', '')

  

    print('test')

    print('man ->', translate('man'))

    print('mans ->', translate('mans'))

    print('king ->', translate('king'))

    print('black ->', translate('black'))

    print('upp ->', translate('upp'))