File size: 8,842 Bytes
2153792
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
from typing import List, Tuple, Union, Dict, Optional, Any, Callable
import numpy as np
import pandas as pd
from collections import Counter


def compute_class_weights(y: Union[List, np.ndarray], method: str = "balanced") -> Union[Dict[int, float], None]:
    if method == "balanced":
        from sklearn.utils.class_weight import compute_class_weight
        classes = np.unique(y)
        weights = compute_class_weight('balanced', classes=classes, y=y)
        return dict(zip(classes, weights))
    else:
        return None


def get_pytorch_weighted_loss(class_weights: Optional[Dict[int, float]] = None,

                              num_classes: Optional[int] = None) -> 'torch.nn.Module':
    try:
        import torch
        import torch.nn as nn
    except ImportError:
        raise ImportError("PyTorch not installed")

    if class_weights is not None:
        weight_tensor = torch.tensor([class_weights[i] for i in sorted(class_weights.keys())], dtype=torch.float)
        return nn.CrossEntropyLoss(weight=weight_tensor)
    else:
        return nn.CrossEntropyLoss()


def get_tensorflow_weighted_loss(class_weights: Optional[Dict[int, float]] = None) -> Callable:
    if not class_weights:
        return 'sparse_categorical_crossentropy'

    weight_list = [class_weights[i] for i in sorted(class_weights.keys())]

    import tensorflow as tf

    def weighted_sparse_categorical_crossentropy(y_true, y_pred):
        y_true = tf.cast(y_true, tf.int32)
        y_true_one_hot = tf.one_hot(y_true, depth=len(weight_list))
        weights = tf.reduce_sum(y_true_one_hot * weight_list, axis=1)
        unweighted_losses = tf.keras.losses.sparse_categorical_crossentropy(y_true, y_pred)
        weighted_losses = unweighted_losses * weights
        return tf.reduce_mean(weighted_losses)

    return weighted_sparse_categorical_crossentropy


def apply_sampling(

        X: np.ndarray,

        y: np.ndarray,

        method: str = "random_under",

        random_state: int = 42

) -> Tuple[np.ndarray, np.ndarray]:
    from imblearn.over_sampling import SMOTE, ADASYN
    from imblearn.under_sampling import RandomUnderSampler
    from imblearn.over_sampling import RandomOverSampler

    if method == "random_under":
        sampler = RandomUnderSampler(random_state=random_state)
    elif method == "random_over":
        sampler = RandomOverSampler(random_state=random_state)
    elif method == "smote":
        sampler = SMOTE(random_state=random_state)
    elif method == "adasyn":
        sampler = ADASYN(random_state=random_state)
    else:
        raise ValueError("method must be one of: random_under, random_over, smote, adasyn")

    X_res, y_res = sampler.fit_resample(X, y)
    return X_res, y_res


def augment_texts(

        texts: List[str],

        labels: List[Any],

        augmentation_type: str = "synonym",

        aug_p: float = 0.1,

        lang: str = "ru",  # language code

        model_name: Optional[str] = None,

        num_aug: int = 1,

        random_state: int = 42

) -> Tuple[List[str], List[Any]]:
    try:
        import nlpaug.augmenter.word as naw
        import nlpaug.augmenter.sentence as nas
    except ImportError:
        raise ImportError("Install nlpaug: pip install nlpaug")

    augmented_texts = []
    augmented_labels = []

    if augmentation_type == "synonym":
        if lang == "en":
            aug = naw.SynonymAug(aug_p=aug_p, aug_max=None)
        else:
            aug = naw.ContextualWordEmbsAug(
                model_path='bert-base-multilingual-cased',
                action="substitute",
                aug_p=aug_p,
                device='cpu'
            )
    elif augmentation_type == "insert":
        aug = naw.RandomWordAug(action="insert", aug_p=aug_p)
    elif augmentation_type == "delete":
        aug = naw.RandomWordAug(action="delete", aug_p=aug_p)
    elif augmentation_type == "swap":
        aug = naw.RandomWordAug(action="swap", aug_p=aug_p)
    elif augmentation_type == "eda":
        aug = naw.AntonymAug()
    elif augmentation_type == "back_trans":
        if not model_name:
            if lang == "ru":
                model_name = "Helsinki-NLP/opus-mt-ru-en"
                back_model = "Helsinki-NLP/opus-mt-en-ru"
            else:
                model_name = "Helsinki-NLP/opus-mt-en-ru"
                back_model = "Helsinki-NLP/opus-mt-ru-en"
        else:
            back_model = model_name

        try:
            from transformers import pipeline
            translator1 = pipeline("translation", model=model_name, tokenizer=model_name)
            translator2 = pipeline("translation", model=back_model, tokenizer=back_model)

            def back_translate(text):
                try:
                    trans = translator1(text)[0]['translation_text']
                    back = translator2(trans)[0]['translation_text']
                    return back
                except Exception:
                    return text

            augmented = [back_translate(t) for t in texts for _ in range(num_aug)]
            labels_aug = [l for l in labels for _ in range(num_aug)]
            return augmented, labels_aug
        except Exception as e:
            print(f"Back-translation failed: {e}. Falling back to synonym augmentation.")
            aug = naw.ContextualWordEmbsAug(model_path='bert-base-multilingual-cased', aug_p=aug_p)
    elif augmentation_type == "llm":
        raise NotImplementedError("LLM-controlled augmentation requires external API (e.g., OpenAI, YandexGPT)")
    else:
        raise ValueError("Unknown augmentation_type")

    for text, label in zip(texts, labels):
        for _ in range(num_aug):
            try:
                aug_text = aug.augment(text)
                if isinstance(aug_text, list):
                    aug_text = aug_text[0]
                augmented_texts.append(aug_text)
                augmented_labels.append(label)
            except Exception as e:
                augmented_texts.append(text)
                augmented_labels.append(label)

    return augmented_texts, augmented_labels


def balance_text_dataset(

        texts: List[str],

        labels: List[Any],

        strategy: str = "augmentation",

        minority_classes: Optional[List[Any]] = None,

        augmentation_type: str = "synonym",

        sampling_method: str = "smote",

        lang: str = "ru",

        embedding_func: Optional[Callable] = None,

        class_weights: bool = False,

        random_state: int = 42

) -> Union[
    Tuple[List[str], List[Any]],  # for augmentation
    Tuple[np.ndarray, np.ndarray, Optional[Dict]]  # for sampling + weights
]:
    label_counts = Counter(labels)
    if minority_classes is None:
        min_count = min(label_counts.values())
        minority_classes = [lbl for lbl, cnt in label_counts.items() if cnt == min_count]

    if strategy == "augmentation":
        minority_texts = [t for t, l in zip(texts, labels) if l in minority_classes]
        minority_labels = [l for l in labels if l in minority_classes]

        aug_texts, aug_labels = augment_texts(
            minority_texts, minority_labels,
            augmentation_type=augmentation_type,
            lang=lang,
            num_aug=max(1, int((max(label_counts.values()) / min_count)) - 1),
            random_state=random_state
        )

        balanced_texts = texts + aug_texts
        balanced_labels = labels + aug_labels
        return balanced_texts, balanced_labels

    elif strategy == "sampling":
        if embedding_func is None:
            raise ValueError("embedding_func is required for sampling strategy")
        X_embed = np.array([embedding_func(t) for t in texts])
        X_res, y_res = apply_sampling(X_embed, np.array(labels), method=sampling_method, random_state=random_state)
        weights = compute_class_weights(y_res) if class_weights else None
        return X_res, y_res, weights

    elif strategy == "both":
        aug_texts, aug_labels = balance_text_dataset(
            texts, labels, strategy="augmentation", minority_classes=minority_classes,
            augmentation_type=augmentation_type, lang=lang, random_state=random_state
        )
        if embedding_func is None:
            return aug_texts, aug_labels
        X_embed = np.array([embedding_func(t) for t in aug_texts])
        X_res, y_res = apply_sampling(X_embed, np.array(aug_labels), method=sampling_method, random_state=random_state)
        weights = compute_class_weights(y_res) if class_weights else None
        return X_res, y_res, weights
    else:
        raise ValueError("strategy must be 'augmentation', 'sampling', or 'both'")