NeuralMachineTranslation/seq2seq.py at master · retkowski/NeuralMachineTranslation · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
import torch
import torch.nn as nn
from torch import optim
import random

from encoder import EncoderRNN
from decoder import DecoderRNN
from LanguageModel import LanguageTokens


class seq2seq():
    def __init__(self, model_config=None, state_dict=None, device=None):
        self.encoder = EncoderRNN(model_config, device=device).to(device)
        self.decoder = DecoderRNN(model_config, device=device).to(device)

        if state_dict:
            self.encoder.load_state_dict(state_dict['encoder'])
            self.decoder.load_state_dict(state_dict['decoder'])

        self.teacher_forcing_ratio = model_config.teacher_forcing_ratio
        self.encoder_optimizer = optim.Adam(
            self.encoder.parameters(),
            lr=model_config.learning_rate)
        self.decoder_optimizer = optim.Adam(
            self.decoder.parameters(),
            lr=model_config.learning_rate)

        if state_dict:
            self.encoder_optimizer.load_state_dict(
                state_dict['encoder_optimizer'])
            self.decoder_optimizer.load_state_dict(
                state_dict['decoder_optimizer'])

        self.criterion = nn.NLLLoss()
        self.beam_width = model_config.beam_width
        self.device = device

    def state_dict(self):
        return {'encoder': self.encoder.state_dict(),
                'decoder': self.decoder.state_dict(),
                'encoder_optimizer': self.encoder_optimizer.state_dict(),
                'decoder_optimizer': self.decoder_optimizer.state_dict()}

    def train(self, input_tensor, target_tensor):
        self.encoder_optimizer.zero_grad()
        self.decoder_optimizer.zero_grad()

        encoder_outputs, decoder_hidden, last_context = self._forward_helper(
            input_tensor)

        decoder_input = self._emptySentenceTensor()
        loss = 0
        use_teacher_forcing = random.random() < self.teacher_forcing_ratio
        output_sentence = []

        for i in range(target_tensor.size(0)):
            decoder_output, decoder_hidden, _, last_context = self.decoder(
                decoder_input, decoder_hidden, encoder_outputs, last_context)
            _, topi = decoder_output.topk(1)
            output_sentence.append(topi.item())

            if use_teacher_forcing:  # Teacher forcing: Feed the target as the next input
                decoder_input = target_tensor[i]  # Teacher forcing
            else:  # Without teacher forcing: use its own predictions as the next input
                decoder_input = topi.squeeze().detach()  # detach from history as input

            loss += self.criterion(decoder_output, target_tensor[i])

            if decoder_input.item() == LanguageTokens.EOS:
                break

        loss.backward()

        self.encoder_optimizer.step()
        self.decoder_optimizer.step()

        return loss.item(), torch.Tensor(output_sentence)

    def predict(self, input_tensor):
        with torch.no_grad():
            encoder_outputs, decoder_hidden, last_context = self._forward_helper(
                input_tensor)

            sequences = [(0.0, [self._emptySentenceTensor()],
                          [], decoder_hidden, [])]

            for _ in range(self.decoder.max_length):
                beam_expansion = []
                for apriori_log_prob, sentence, decoder_outputs, decoder_hidden, attention_weights_list in sequences:
                    decoder_input = sentence[-1]
                    if(decoder_input.item() != LanguageTokens.EOS):
                        decoder_output, decoder_hidden, attention_weights, last_context = self.decoder(
                            decoder_input, decoder_hidden, encoder_outputs, last_context)

                        log_probabilities, indexes = decoder_output.squeeze().data.topk(
                            self.beam_width)

                        for i in range(len(log_probabilities)):
                            log_prob = log_probabilities[i]
                            index = indexes[i]

                            beam_expansion.append(
                                (apriori_log_prob + log_prob,
                                 sentence + [index],
                                    decoder_outputs + [decoder_output],
                                    decoder_hidden,
                                    attention_weights_list + [attention_weights]))
                    else:
                        beam_expansion.append(
                            (apriori_log_prob,
                             sentence,
                             decoder_outputs,
                             decoder_hidden,
                             attention_weights_list))

                sequences = sorted(
                    beam_expansion,
                    reverse=True,
                    key=lambda x: x[0])[
                    :self.beam_width]

            # best sequence
            _, sentence, decoder_outputs, _, attention_weights = sequences[0]
            return torch.tensor(sentence), decoder_outputs, attention_weights

    def evaluate(self, input_tensor, target_tensor):
        with torch.no_grad():
            sentence, decoder_outputs, _ = self.predict(input_tensor)

            target_length = target_tensor.size(0)

            loss = sum([self.criterion(decoder_outputs[i], target_tensor[i])
                        for i in range(min(len(decoder_outputs), target_length))])

            return loss.item(), sentence

    def _forward_helper(self, input_tensor):
        encoder_hidden = self.encoder.initEncoderHidden()

        encoder_outputs, encoder_hidden = self.encoder(
            input_tensor, encoder_hidden)
        # remove batch dimension
        encoder_outputs = encoder_outputs.view(-1, self.decoder.hidden_size)
        decoder_hidden = self.decoder.getHidden(encoder_hidden)
        last_context = self.decoder.initContext()

        return encoder_outputs, decoder_hidden, last_context

    def _emptySentenceTensor(self):
        return torch.tensor([[LanguageTokens.SOS]], device=self.device)