Esempio n. 1
0
 def test_sanity(self):
     data = LanguageData(['how are you?', 'I like it', 'cereja', 'cherry'],
                         to_lower=True)
     self.assertEqual(
         data.synergy(['how are you?', 'I like it', 'cereja', 'cherry']), 1)
     self.assertEqual(data.synergy(['how are you?', 'I like it', 'cereja']),
                      0.889)
Esempio n. 2
0
 def __init__(self,
              source_data,
              target_data,
              source_name=None,
              target_name=None,
              **kwargs):
     self.source = LanguageData(source_data, name=source_name, **kwargs)
     self.target = LanguageData(target_data, name=target_name, **kwargs)
     self._percent_train = 0.8
     self._n_train = self.n_train
     self._valid_parallel_data(self.source.data, self.target.data)
Esempio n. 3
0
 def __init__(self, source_data, target_data, source_name=None, percent_train=0.8, target_name=None, stop_words=(),
              punctuation='!?,.', to_lower=True, is_remove_punctuation=True, is_remove_stop_words=True,
              is_remove_accent=False, is_destructive=False):
     self.source = LanguageData(source_data, name=source_name, stop_words=stop_words,
                                punctuation=punctuation, to_lower=to_lower,
                                is_remove_punctuation=is_remove_punctuation,
                                is_remove_stop_words=is_remove_stop_words,
                                is_remove_accent=is_remove_accent, is_destructive=is_destructive)
     self.target = LanguageData(target_data, name=target_name, stop_words=stop_words,
                                punctuation=punctuation, to_lower=to_lower,
                                is_remove_punctuation=is_remove_punctuation,
                                is_remove_stop_words=is_remove_stop_words,
                                is_remove_accent=is_remove_accent, is_destructive=is_destructive)
     self._percent_train = percent_train
     self._n_train = self.n_train
     self._valid_parallel_data(self.source.data, self.target.data)
Esempio n. 4
0
class Corpus(object):
    def __init__(self,
                 source_data,
                 target_data,
                 source_name=None,
                 target_name=None,
                 **kwargs):
        self.source = LanguageData(source_data, name=source_name, **kwargs)
        self.target = LanguageData(target_data, name=target_name, **kwargs)
        self._percent_train = 0.8
        self._n_train = self.n_train
        self._valid_parallel_data(self.source.data, self.target.data)

    def __iter__(self):
        return zip(self.source, self.target)

    def __len__(self):
        return len(self.source)

    def __repr__(self):
        return f'Corpus(examples: {len(self)} - source_vocab_size: {self.source.vocab_size} - target_vocab_size:{self.target.vocab_size})'

    def __getitem__(self, item):
        if isinstance(item, int):
            return self.source.data[item], self.target.data[item]
        return list(zip(self.source.data[item], self.target.data[item]))

    @property
    def source_language(self):
        """
        Only read. The setter function is `Corpus.source.config.set_config(name="new name")`
        """
        return self.source.config.name

    @property
    def target_language(self):
        """
        Only read. The setter function is `Corpus.target.config.set_config(name="new name")`
        """
        return self.target.config.name

    @property
    def config(self):
        """
        Only read. The setter function is `Corpus.<source or target>.config.set_config`
        """
        return {
            'source': self.source.config.get(),
            'target': self.source.config.get()
        }

    def set_config(self, **kwargs):
        """
        you can use `Corpus.config` which returns a dictionary with the current configuration.
        And then just use the this method `Corpus.set_config` by sending the changed dictionary.

        :param kwargs: dictionary like this:
                    {
                     'source': {'name': 'SOURCE_NAME',
                                'punctuation': '',
                                'remove_punctuation': False,
                                'remove_stop_words': False,
                                'stop_words': (),
                                'to_lower': False
                                },
                     'target': {'name':'TARGET_NAME',
                                'punctuation': '',
                                'remove_punctuation': False,
                                'remove_stop_words': False,
                                'stop_words': (),
                                'to_lower': False
                                }
                    }
        """
        source_config = kwargs.get('source')
        target_config = kwargs.get('target')
        if source_config:
            self.source.config.set_config(**source_config)
        if target_config:
            self.target.config.set_config(**target_config)

    @staticmethod
    def is_parallel(data):
        try:
            for x, y in data:
                if isinstance(x, (str, int, float)) and isinstance(
                        y, (str, int, float)):
                    return True
                break
        except:
            pass
        return False

    @classmethod
    def distinct_from_parallel(cls, data):
        return get_cols(data)

    @classmethod
    def load_from_parallel_data(cls,
                                data,
                                source_name: str = None,
                                target_name: str = None,
                                **kwargs):
        if cls.is_parallel(data):
            source_data, target_data = cls.distinct_from_parallel(data)
            return cls(source_data,
                       target_data,
                       source_name=source_name,
                       target_name=target_name,
                       **kwargs)
        raise ValueError("isn't valid parallel data")

    @property
    def n_train(self):
        return int(self._percent_train * len(self.source.data))

    @property
    def n_test(self):
        return len(self.source.data) - self.n_train

    def _can_go_test(self, x, y):
        x = self.source.preprocess(x)
        y = self.target.preprocess(y)
        if self.source.phrases_freq.get(x) == 1 and len(x.split()) >= 4:
            x, y = x.split(), y.split()
            for i in x:
                if self.source.words_freq.get(i) <= x.count(i):
                    return False
            for i in y:
                if self.target.words_freq.get(i) <= y.count(i):
                    return False
            return True
        return False

    def _valid_parallel_data(self, x, y):
        assert len(x) == len(
            y
        ), f"Size of {self.source_language} ({len(x)}) != {self.target_language} ({len(y)})"

    def _update_filters(self, x, y):
        x = self.source.preprocess(x)
        y = self.target.preprocess(y)
        for i in x.split():
            self.source.words_freq.subtract([i])

        for i in y.split():
            self.target.words_freq.subtract([i])

        self.source.phrases_freq.subtract([x])
        self.target.phrases_freq.subtract([y])

    def _get_vocab_data(self,
                        source_vocab_size: int = None,
                        target_vocab_size: int = None,
                        order='most_common'):
        source_vocab_data = {}
        target_vocab_data = {}
        if source_vocab_size is not None:
            source_vocab_data = self.source.sample_words_freq(
                max_items=source_vocab_size, order=order)
        if target_vocab_size is not None:
            target_vocab_data = self.target.sample_words_freq(
                max_items=target_vocab_size, order=order)

        for x, y in zip(self.source.data, self.target.data):
            if source_vocab_size:
                if not all(
                        list(
                            map(lambda w: w in source_vocab_data,
                                self.source.preprocess(x).split()))):
                    continue
            if target_vocab_size:
                if not all(
                        list(
                            map(lambda w: w in target_vocab_data,
                                self.target.preprocess(y).split()))):
                    continue
            yield [x, y]

    def save(self,
             save_on_dir: str,
             take_split: bool = True,
             test_max_size: int = None,
             source_vocab_size: int = None,
             target_vocab_size: int = None,
             shuffle=True,
             prefix=None,
             ext='align',
             **kwargs):
        save_on_dir = Path(save_on_dir)
        if take_split:
            x_train, y_train, x_test, y_test = self.split_data(
                test_max_size=test_max_size,
                source_vocab_size=source_vocab_size,
                target_vocab_size=target_vocab_size,
                take_parallel_data=False,
                shuffle=shuffle)
            train_prefix, test_prefix = (
                f'{prefix}_train',
                f'{prefix}_test') if prefix is not None else ('train', 'test')
            data_to_save = ((train_prefix, x_train, y_train), (test_prefix,
                                                               x_test, y_test))
        else:
            data_to_save = ((prefix, self.source.data, self.target.data), )

        for prefix, x, y in data_to_save:
            save_on = save_on_dir.join(
                f'{prefix}_{self.source_language}.{ext.strip(".")}')
            FileIO.create(save_on, data=x).save(**kwargs)
            save_on = save_on_dir.join(
                f'{prefix}_{self.target_language}.{ext.strip(".")}')
            FileIO.create(save_on, data=y).save(**kwargs)

    @classmethod
    def load_corpus_from_csv(cls,
                             path_: str,
                             src_col_name: str,
                             trg_col_name: str,
                             source_name=None,
                             target_name=None):

        csv_read = csv.DictReader(FileIO.load(path_).data)
        src_data = []
        trg_data = []
        for i in csv_read:
            for col_name in (src_col_name, trg_col_name):
                if col_name not in i:
                    raise ValueError(
                        f"Not found col <{col_name}> in {list(i.keys())}")
            src_data.append(i[src_col_name])
            trg_data.append(i[trg_col_name])
        return cls(src_data,
                   trg_data,
                   source_name=source_name,
                   target_name=target_name)

    def split_data(self,
                   test_max_size: int = None,
                   source_vocab_size: int = None,
                   target_vocab_size: int = None,
                   shuffle=True,
                   take_parallel_data=True,
                   take_corpus_instances=False,
                   legacy_test=None):
        """
        Guarantees test data without data identical to training and only with vocabulary that exists in training


        :param test_max_size: int = max examples on test data
        :param source_vocab_size: int = restrict most common vocab
        :param target_vocab_size: int = restrict most common vocab
        :param shuffle: bool = randomize
        :param take_parallel_data: bool = zipped data if true else return (x_train, y_train, x_test, y_test)
        :param take_corpus_instances: bool = return new instances for train data and test data
        :param legacy_test: List[Tuple[str,str]] = parallel data
        """
        self.source.reset_freq()
        self.target.reset_freq()
        train = []
        test = []

        if legacy_test is not None:
            test = Corpus(*self.distinct_from_parallel(legacy_test),
                          source_name=self.source_language,
                          target_name=self.target_language)

        test_max_size = test_max_size if test_max_size is not None and isinstance(
            test_max_size,
            (int, float)) else len(self.source.data) - self.n_train
        if source_vocab_size is not None or target_vocab_size is not None:
            data = list(
                self._get_vocab_data(source_vocab_size=source_vocab_size,
                                     target_vocab_size=target_vocab_size))
        else:
            data = list(zip(self.source.data, self.target.data))

        if shuffle:
            random.shuffle(data)

        for x, y in data:
            # remove blank line
            if x == '' or y == '':
                continue
            if legacy_test is not None:
                # remove sentence from train.
                if self.source.preprocess(x) in test.source.phrases_freq:
                    continue
            if (self._can_go_test(x, y)
                    and len(test) < test_max_size) and legacy_test is None:
                test.append([x, y])
                self._update_filters(x, y)
                continue
            train.append([x, y])

        if take_parallel_data is False:
            return (*get_cols(train), *get_cols(test))
        if take_corpus_instances is True:
            train = self.load_from_parallel_data(train, self.source_language,
                                                 self.target_language)
            test = self.load_from_parallel_data(test, self.source_language,
                                                self.target_language)
            return train, test
        return train, test

    def split_data_and_save(self, **kwargs):
        alternative = f"You can use <Corpus.save>"
        warnings.warn(
            f"This function has been deprecated and will be removed in future versions. "
            f"{alternative}", DeprecationWarning, 2)
        self.save(**kwargs)