NeuralMachineTranslation/main.py at master · gugundi/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
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
from bleu import compute_bleu
from device import select_device, with_cpu, with_gpu
import json
from parse import get_config
from random import sample
from tensorboardX import SummaryWriter
import torchtext
import torch
from torch.nn.utils import clip_grad_norm_
from utils import filter_words, get_or_create_dir, get_text, list2words, torch2words
from visualize import visualize_attention


def main():
    use_gpu, device, device_idx = select_device()
    if use_gpu:
        device_name = torch.cuda.get_device_name(device_idx)
        print(f'Using device: {device} ({device_name})')
        with torch.cuda.device(device_idx):
            run(use_gpu, device, device_idx)
    else:
        print(f'Using device: cpu')
        run(use_gpu, device, device_idx)


def run(use_gpu, device, device_idx):
    config = get_config(use_gpu, device, device_idx)
    val_iter = config.get('val_iter')
    val_data = val_iter.data()
    val_dataset = config.get('val_dataset')

    def sample_validation_batches(k):
        return torchtext.data.Batch(sample(val_data, k), val_dataset, device)

    # save source and target language vocabularies
    source_language = config.get('src_language')
    target_language = config.get('trg_language')
    data = {
        "source": {
            "itos": source_language.itos,
            "stoi": source_language.stoi,
        },
        "target": {
            "itos": target_language.itos,
            "stoi": target_language.stoi,
        },
    }
    weights_path = config.get('weights_path')
    with open(f'{weights_path}/language.json', 'w') as f:
        json.dump(data, f)

    train(config, sample_validation_batches)


def train(config, sample_validation_batches):
    source_language = config.get('src_language')
    target_language = config.get('trg_language')
    EOS_token = config.get('EOS_token')
    PAD_token = config.get('PAD_token')
    SOS_token = config.get('SOS_token')
    train_iter = config.get('train_iter')
    val_iter = config.get('val_iter')
    writer_path = config.get('writer_path')
    writer_train_path = get_or_create_dir(writer_path, 'train')
    writer_val_path = get_or_create_dir(writer_path, 'val')
    writer_train = SummaryWriter(log_dir=writer_train_path)
    writer_val = SummaryWriter(log_dir=writer_val_path)
    epochs = config.get('epochs')
    training = config.get('training')
    eval_every = training.get('eval_every')
    sample_every = training.get('sample_every')
    use_attention = config.get('use_attention')
    step = 1
    for epoch in range(epochs):
        print(f'Epoch: {epoch+1}/{epochs}')
        save_weights(config)
        for i, training_batch in enumerate(train_iter):
            loss = train_batch(config, training_batch)
            writer_train.add_scalar('loss', loss, step)

            if step == 1 or step % eval_every == 0:
                val_lengths = 0
                val_losses = 0
                reference_corpus = []
                translation_corpus = []
                for val_batch in val_iter:
                    val_loss, translations = evaluate_batch(config, val_batch)
                    val_lengths += 1
                    val_losses += val_loss
                    val_batch_trg, _ = val_batch.trg
                    _, batch_size = val_batch_trg.shape
                    references = map(lambda i: torch2words(target_language, val_batch_trg[:, i]), range(batch_size))
                    references = map(lambda words: [list(filter_words(words, SOS_token, EOS_token, PAD_token))], references)
                    reference_corpus.extend(references)
                    translations = map(lambda translation: list2words(target_language, translation), translations)
                    translations = map(lambda words: list(filter_words(words, SOS_token, EOS_token, PAD_token)), translations)
                    translation_corpus.extend(translations)
                bleu = compute_bleu(reference_corpus, translation_corpus)
                val_loss = val_losses / val_lengths
                writer_val.add_scalar('bleu', bleu, step)
                writer_val.add_scalar('loss', val_loss, step)

            if step % sample_every == 0:
                val_batch = sample_validation_batches(1)
                val_batch_src, val_lengths_src = val_batch.src
                val_batch_trg, _ = val_batch.trg
                s0 = val_lengths_src[0].item()
                _, translations, attention_weights = evaluate_batch(config, val_batch, True)
                source_words = torch2words(source_language, val_batch_src[:, 0])
                target_words = torch2words(target_language, val_batch_trg[:, 0])
                translation_words = list(filter(lambda word: word != PAD_token, list2words(target_language, translations[0])))
                if use_attention and sum(attention_weights.shape) != 0:
                    attention_figure = visualize_attention(source_words[:s0], translation_words, with_cpu(attention_weights))
                    writer_val.add_figure('attention', attention_figure, step)
                text = get_text(source_words, target_words, translation_words, SOS_token, EOS_token, PAD_token)
                writer_val.add_text('translation', text, step)

            step += 1

    save_weights(config)


def train_batch(config, batch):
    model = config.get('model')
    optimizer = config.get('optimizer')
    gradient_clipping = config.get('gradient_clipping')

    model.train()
    ys = model(batch)
    loss = get_loss(config, batch, ys)

    optimizer.zero_grad()
    loss.backward()
    if gradient_clipping:
        clip_grad_norm_(model.parameters(), 1)
    optimizer.step()

    return with_cpu(loss)


def evaluate_batch(config, batch, sample=False):
    model = config.get('model')

    with torch.no_grad():
        model.eval()
        if sample:
            ys, translations, attention_weights = model(batch, training=False, sample=True)
            loss = get_loss(config, batch, ys)
            return with_cpu(loss), translations, attention_weights
        else:
            ys, translations = model(batch, training=False, sample=False)
            loss = get_loss(config, batch, ys)
            return with_cpu(loss), translations


def create_mask(batch_tuple):
    batch, lengths = batch_tuple
    max_length, batch_size = batch.shape
    mask = with_gpu(torch.ones(max_length, batch_size))
    for i, length in enumerate(lengths):
        mask[length:, i] = 0
    return mask


def compute_batch_loss(loss, mask, lengths):
    loss = loss * mask
    loss = torch.sum(loss, 0, dtype=torch.float)
    loss = loss / lengths.float()
    loss = loss.mean()
    return loss


def get_loss(config, batch, ys):
    loss_fn = config.get('loss_fn')
    mask = create_mask(batch.trg)
    target_batch, target_lengths = batch.trg
    T, batch_size = target_batch.shape
    losses = with_gpu(torch.empty((T, batch_size), dtype=torch.float))
    for i in range(T):
        losses[i] = loss_fn(ys[i], target_batch[i])
    loss = compute_batch_loss(losses, mask, target_lengths)
    return loss


def save_weights(config):
    weights_path = config.get("weights_path")
    model_path = f'{weights_path}/model'
    model = config.get('model')
    model_weights = model.state_dict()
    torch.save(model_weights, model_path)


if __name__ == '__main__':
    main()