transformer/data.py at new-branch · thoongee/transformer · 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
import os
import torch
from torch.utils.data import DataLoader
from torchtext.vocab import build_vocab_from_iterator
import torchtext.transforms as T

from utils import save_pkl, load_pkl


class Multi30k():

    def __init__(self,
                 lang=("en", "de"),
                 max_seq_len=256,
                 unk_idx=0,
                 pad_idx=1,
                 sos_idx=2,
                 eos_idx=3,
                 vocab_min_freq=2):

        self.dataset_name = "multi30k"
        self.lang_src, self.lang_tgt = lang
        self.max_seq_len = max_seq_len
        self.unk_idx = unk_idx
        self.pad_idx = pad_idx
        self.sos_idx = sos_idx
        self.eos_idx = eos_idx
        self.unk = "<unk>"
        self.pad = "<pad>"
        self.sos = "<sos>"
        self.eos = "<eos>"
        self.specials={
                self.unk: self.unk_idx,
                self.pad: self.pad_idx,
                self.sos: self.sos_idx,
                self.eos: self.eos_idx
                }
        self.vocab_min_freq = vocab_min_freq
        # print('self.lang_src: ',self.lang_src) # en
        # print('self.lang_tgt: ',self.lang_tgt) # de

        # Build tokenizers for source and target languages
        self.tokenizer_src = self.build_tokenizer(self.lang_src)
        self.tokenizer_tgt = self.build_tokenizer(self.lang_tgt)

        # Initialize dataset and vocab variables
        self.train = None
        self.valid = None
        self.test = None
        self.build_dataset()

        self.vocab_src = None
        self.vocab_tgt = None
        self.build_vocab()

        self.transform_src = None
        self.transform_tgt = None
        self.build_transform()


    def build_dataset(self, raw_dir="raw", cache_dir=".data"):
        # Build dataset from raw files or load from cache
        cache_dir = os.path.join(cache_dir, self.dataset_name)
        raw_dir = os.path.join(cache_dir, raw_dir)
        os.makedirs(raw_dir, exist_ok=True)

        train_file = os.path.join(cache_dir, "train.pkl")
        valid_file = os.path.join(cache_dir, "valid.pkl")
        test_file = os.path.join(cache_dir, "test.pkl")

        if os.path.exists(train_file):
            self.train = load_pkl(train_file)
        else:
            with open(os.path.join(raw_dir, "train.en"), "r") as f:
                train_en = [text.rstrip() for text in f]
            with open(os.path.join(raw_dir, "train.de"), "r") as f:
                train_de = [text.rstrip() for text in f]
            self.train = [(en, de) for en, de in zip(train_en, train_de)]
            save_pkl(self.train , train_file)

        if os.path.exists(valid_file):
            self.valid = load_pkl(valid_file)
        else:
            with open(os.path.join(raw_dir, "val.en"), "r") as f:
                valid_en = [text.rstrip() for text in f]
            with open(os.path.join(raw_dir, "val.de"), "r") as f:
                valid_de = [text.rstrip() for text in f]
            self.valid = [(en, de) for en, de in zip(valid_en, valid_de)]
            save_pkl(self.valid, valid_file)

        if os.path.exists(test_file):
            self.test = load_pkl(test_file)
        else:
            with open(os.path.join(raw_dir, "test_2016_flickr.en"), "r") as f:
                test_en = [text.rstrip() for text in f]
            with open(os.path.join(raw_dir, "test_2016_flickr.de"), "r") as f:
                test_de = [text.rstrip() for text in f]
            self.test = [(en, de) for en, de in zip(test_en, test_de)]
            save_pkl(self.test, test_file)


    def build_vocab(self, cache_dir=".data"):
        # Build vocabulary from training data or load from cache
        assert self.train is not None
        def yield_tokens(is_src=True):
            for text_pair in self.train:
                if is_src:
                    yield [str(token) for token in self.tokenizer_src(text_pair[0])]
                else:
                    yield [str(token) for token in self.tokenizer_tgt(text_pair[1])]

        cache_dir = os.path.join(cache_dir, self.dataset_name)
        os.makedirs(cache_dir, exist_ok=True)

        vocab_src_file = os.path.join(cache_dir, f"vocab_{self.lang_src}.pkl")
        if os.path.exists(vocab_src_file):
            vocab_src = load_pkl(vocab_src_file)
        else:
            vocab_src = build_vocab_from_iterator(yield_tokens(is_src=True), min_freq=self.vocab_min_freq, specials=self.specials.keys())
            vocab_src.set_default_index(self.unk_idx)
            save_pkl(vocab_src, vocab_src_file)

        vocab_tgt_file = os.path.join(cache_dir, f"vocab_{self.lang_tgt}.pkl")
        if os.path.exists(vocab_tgt_file):
            vocab_tgt = load_pkl(vocab_tgt_file)
        else:
            vocab_tgt = build_vocab_from_iterator(yield_tokens(is_src=False), min_freq=self.vocab_min_freq, specials=self.specials.keys())
            vocab_tgt.set_default_index(self.unk_idx)
            save_pkl(vocab_tgt, vocab_tgt_file)

        self.vocab_src = vocab_src
        self.vocab_tgt = vocab_tgt


    def build_tokenizer(self, lang):
        from torchtext.data.utils import get_tokenizer
        spacy_lang_dict = {
                'en': "en_core_web_sm",
                'de': "de_core_news_sm"
                }
        assert lang in spacy_lang_dict.keys()
        # print('spacy_lang_dict[lang]: ',spacy_lang_dict[lang]) # en_core_web_sm

        # Return tokenizer for the specified language
        return get_tokenizer("spacy", spacy_lang_dict[lang])


    def build_transform(self):
        # Build transform functions for source and target sequences
        def get_transform(self, vocab):
            return T.Sequential(
                    T.VocabTransform(vocab),
                    T.Truncate(self.max_seq_len-2),
                    T.AddToken(token=self.sos_idx, begin=True),
                    T.AddToken(token=self.eos_idx, begin=False),
                    T.ToTensor(padding_value=self.pad_idx))

        self.transform_src = get_transform(self, self.vocab_src)
        self.transform_tgt = get_transform(self, self.vocab_tgt)


    def collate_fn(self, pairs):
        # Collate function to process a batch of data pairs
        src = [self.tokenizer_src(pair[0]) for pair in pairs]
        tgt = [self.tokenizer_tgt(pair[1]) for pair in pairs]
        batch_src = self.transform_src(src)
        batch_tgt = self.transform_tgt(tgt)
        return (batch_src, batch_tgt)


    def get_iter(self, **kwargs):
        # Get data loaders for train, valid, and test sets
        if self.transform_src is None:
            self.build_transform()
        train_iter = DataLoader(self.train, collate_fn=self.collate_fn, **kwargs)
        valid_iter = DataLoader(self.valid, collate_fn=self.collate_fn, **kwargs)
        test_iter = DataLoader(self.test, collate_fn=self.collate_fn, **kwargs)
        return train_iter, valid_iter, test_iter


    def translate(self, model, src_sentence: str, decode_func):
        # Translate a source sentence using the provided model and decode function
        model.eval()
        src = self.transform_src([self.tokenizer_src(src_sentence)]).view(1, -1)
        num_tokens = src.shape[1]
        tgt_tokens = decode_func(model,
                                 src,
                                 max_len=num_tokens+5,
                                 start_symbol=self.sos_idx,
                                 end_symbol=self.eos_idx).flatten().cpu().numpy()
        tgt_sentence = " ".join(self.vocab_tgt.lookup_tokens(tgt_tokens))
        return tgt_sentence