File size: 11,717 Bytes
48b5e1d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
#!/usr/bin/env python

import os
import torch
import string
import onnxruntime as ort
from dataclasses import dataclass
from omegaconf import OmegaConf
from typing import List, Optional, Union, Dict
from sentencepiece import SentencePieceProcessor
from torch.utils.data import Dataset, DataLoader
from typing import Iterator, List, Iterable, Tuple

ACRONYM_TOKEN = "<ACRONYM>"
torch.set_grad_enabled(False)
torch.backends.cudnn.enabled = False
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"


@dataclass
class PunctCapConfigONNX:
    spe_filename: str = "xlm_roberta_encoding.model"
    model_filename: str = "nemo_model.onnx"
    config_filename: str = "config.yaml"
    directory: Optional[str] = None


class PunctCapModelONNX:
    def __init__(self, cfg: PunctCapConfigONNX):
        self._spe_path = os.path.join(cfg.directory, cfg.spe_filename)
        onnx_path = os.path.join(cfg.directory, cfg.model_filename)
        config_path = os.path.join(cfg.directory, cfg.config_filename)

        self._tokenizer: SentencePieceProcessor = SentencePieceProcessor(self._spe_path)
        self._ort_session: ort.InferenceSession = ort.InferenceSession(onnx_path)
        self._config = OmegaConf.load(config_path)
        self._max_len = self._config.max_length
        self._pre_labels: List[str] = self._config.pre_labels
        self._post_labels: List[str] = self._config.post_labels
        self._languages: List[str] = self._config.languages
        self._null_token = self._config.get("null_token", "<NULL>")

    def _setup_dataloader(self, texts: List[str], batch_size_tokens: int, overlap: int) -> DataLoader:
        dataset: TextInferenceDataset = TextInferenceDataset(
            texts=texts,
            batch_size_tokens=batch_size_tokens,
            overlap=overlap,
            max_length=self._max_len,
            spe_model_path=self._spe_path,
        )
        return DataLoader(
            dataset=dataset,
            collate_fn=dataset.collate_fn,
            batch_sampler=dataset.sampler,
        )

    def punctuation_removal(self, texts: List[str]) -> List[str]:
        punkt = string.punctuation + """`÷×؛<>_()*&^%][ـ،/:"؟.,'{}~¦+|!”…–ـ""" + """!?。。"""
        punkt = punkt.replace("-", "")
        punkt = punkt.replace("'", "")
        punkt += "„“"
        return [text.translate(str.maketrans("", "", punkt)).lower().strip() for text in texts]

    def infer(
        self,
        texts: List[str],
        apply_sbd: bool = False,
        batch_size_tokens: int = 4096,
        overlap: int = 16,
    ) -> Union[List[str], List[List[str]]]:
        texts = self.punctuation_removal(texts)

        collectors: List[PunctCapCollector] = [
            PunctCapCollector(sp_model=self._tokenizer, apply_sbd=apply_sbd, overlap=overlap)
            for _ in range(len(texts))
        ]
        dataloader: DataLoader = self._setup_dataloader(texts=texts, batch_size_tokens=batch_size_tokens, overlap=overlap)
        for batch in dataloader:
            input_ids, batch_indices, input_indices, lengths = batch
            pre_preds, post_preds, cap_preds, seg_preds = self._ort_session.run(None, {"input_ids": input_ids.numpy()})
            batch_size = input_ids.shape[0]
            for i in range(batch_size):
                length = lengths[i].item()
                batch_idx = batch_indices[i].item()
                input_idx = input_indices[i].item()
                segment_ids = input_ids[i, 1 : length - 1].tolist()
                segment_pre_preds = pre_preds[i, 1 : length - 1].tolist()
                segment_post_preds = post_preds[i, 1 : length - 1].tolist()
                segment_cap_preds = cap_preds[i, 1 : length - 1].tolist()
                segment_sbd_preds = seg_preds[i, 1 : length - 1].tolist()
                pre_tokens = [self._pre_labels[i] for i in segment_pre_preds]
                post_tokens = [self._post_labels[i] for i in segment_post_preds]
                pre_tokens = [x if x != self._null_token else None for x in pre_tokens]
                post_tokens = [x if x != self._null_token else None for x in post_tokens]
                collectors[batch_idx].collect(
                    ids=segment_ids,
                    pre_preds=pre_tokens,
                    post_preds=post_tokens,
                    cap_preds=segment_cap_preds,
                    sbd_preds=segment_sbd_preds,
                    idx=input_idx,
                )
        outputs: Union[List[str], List[List[str]]] = [x.produce() for x in collectors]
        return outputs


@dataclass
class TokenizedSegment:
    input_ids: List[int]
    batch_idx: int
    input_idx: int

    def __len__(self) -> int:
        return len(self.input_ids)


class TokenBatchSampler(Iterable):
    def __init__(self, segments: List[TokenizedSegment], batch_size_tokens: int):
        self._batches = self._make_batches(segments, batch_size_tokens)

    def _make_batches(self, segments: List[TokenizedSegment], batch_size_tokens: int) -> List[List[int]]:
        segments_with_index = [(segment, i) for i, segment in enumerate(segments)]
        segments_with_index.sort(key=lambda x: len(x[0]), reverse=True)

        batches, current_batch_elements, current_max_len = [], [], 0

        for segment, idx in segments_with_index:
            potential_max_len = max(current_max_len, len(segment))

            if potential_max_len * (len(current_batch_elements) + 1) > batch_size_tokens:
                batches.append(current_batch_elements)
                current_batch_elements, current_max_len = [], 0

            current_batch_elements.append(idx)
            current_max_len = potential_max_len

        if current_batch_elements:
            batches.append(current_batch_elements)

        return batches

    def __iter__(self) -> Iterator:
        yield from self._batches

    def __len__(self) -> int:
        return len(self._batches)


class TextInferenceDataset(Dataset):
    def __init__(
        self,
        texts: List[str],
        spe_model_path: str,
        batch_size_tokens: int = 4096,
        max_length: int = 512,
        overlap: int = 32,
    ):
        self._spe_model = SentencePieceProcessor(spe_model_path)
        self._segments = self._tokenize_inputs(texts, max_length, overlap)
        self._sampler = TokenBatchSampler(self._segments, batch_size_tokens)

    @property
    def sampler(self) -> Iterable:
        return self._sampler

    def _tokenize_inputs(self, texts: List[str], max_len: int, overlap: int) -> List[TokenizedSegment]:
        max_len -= 2
        segments = []

        for batch_idx, text in enumerate(texts):
            ids, start, input_idx = self._spe_model.EncodeAsIds(text), 0, 0

            while start < len(ids):
                adjusted_start = start - overlap if input_idx else 0
                segments.append(
                    TokenizedSegment(
                        ids[adjusted_start : adjusted_start + max_len],
                        batch_idx,
                        input_idx,
                    )
                )
                start += max_len - overlap
                input_idx += 1

        return segments

    def __len__(self) -> int:
        return len(self._segments)

    def __getitem__(self, idx: int) -> Tuple[torch.Tensor, int, int]:
        segment = self._segments[idx]
        input_ids = torch.Tensor([self._spe_model.bos_id(), *segment.input_ids, self._spe_model.eos_id()])
        return input_ids, segment.batch_idx, segment.input_idx

    def collate_fn(self, batch: List[Tuple[torch.Tensor, int, int]]) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
        input_ids = [x[0] for x in batch]
        lengths = torch.tensor([x.shape[0] for x in input_ids])
        max_len = lengths.max().item()

        batched_ids = torch.full((len(input_ids), max_len), self._spe_model.pad_id())
        for idx, ids in enumerate(input_ids):
            batched_ids[idx, : lengths[idx]] = ids

        return (
            batched_ids,
            torch.tensor([x[1] for x in batch]),
            torch.tensor([x[2] for x in batch]),
            lengths,
        )


@dataclass
class PCSegment:
    ids: List[int]
    pre_preds: List[Optional[str]]
    post_preds: List[Optional[str]]
    cap_preds: List[List[int]]
    sbd_preds: List[int]

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


class PunctCapCollector:
    def __init__(self, apply_sbd: bool, overlap: int, sp_model: SentencePieceProcessor):
        self._segments: Dict[int, PCSegment] = {}
        self._apply_sbd = apply_sbd
        self._overlap = overlap
        self._sp_model = sp_model

    def collect(
        self,
        ids: List[int],
        pre_preds: List[Optional[str]],
        post_preds: List[Optional[str]],
        sbd_preds: List[int],
        cap_preds: List[List[int]],
        idx: int,
    ):
        self._segments[idx] = PCSegment(
            ids=ids,
            pre_preds=pre_preds,
            post_preds=post_preds,
            sbd_preds=sbd_preds,
            cap_preds=cap_preds,
        )

    def produce(self) -> Union[List[str], str]:
        ids: List[int] = []
        pre_preds: List[Optional[str]] = []
        post_preds: List[Optional[str]] = []
        cap_preds: List[List[int]] = []
        sbd_preds: List[int] = []

        for i in range(len(self._segments)):
            segment = self._segments[i]
            start = 0
            stop = len(segment)
            if i > 0:
                start += self._overlap // 2
            if i < len(self._segments) - 1:
                stop -= self._overlap // 2

            ids.extend(segment.ids[start:stop])
            pre_preds.extend(segment.pre_preds[start:stop])
            post_preds.extend(segment.post_preds[start:stop])
            sbd_preds.extend(segment.sbd_preds[start:stop])
            cap_preds.extend(segment.cap_preds[start:stop])

        input_tokens = [self._sp_model.IdToPiece(x) for x in ids]
        output_texts: List[str] = []
        current_chars: List[str] = []

        for token_idx, token in enumerate(input_tokens):
            if token.startswith("▁") and current_chars:
                current_chars.append(" ")
            char_start = 1 if token.startswith("▁") else 0

            for token_char_idx, char in enumerate(token[char_start:], start=char_start):
                if token_char_idx == char_start and pre_preds[token_idx] is not None:
                    current_chars.append(pre_preds[token_idx])
                if cap_preds[token_idx][token_char_idx]:
                    char = char.upper()
                current_chars.append(char)

                label = post_preds[token_idx]
                if label == ACRONYM_TOKEN:
                    current_chars.append(".")
                elif token_char_idx == len(token) - 1 and post_preds[token_idx] is not None:
                    current_chars.append(post_preds[token_idx])
                if self._apply_sbd and token_char_idx == len(token) - 1 and sbd_preds[token_idx]:
                    output_texts.append("".join(current_chars))
                    current_chars = []

        if current_chars:
            output_texts.append("".join(current_chars))
        if not self._apply_sbd:
            if len(output_texts) > 1:
                raise ValueError(f"Not applying SBD but got more than one result: {output_texts}")
            return output_texts[0]
        return output_texts


class MultiLingual:
    def __init__(self):
        cfg = PunctCapConfigONNX(directory="/code/models/multilingual")
        self._punctuator = PunctCapModelONNX(cfg)

    def punctuate(self, data: str) -> str:
        return self._punctuator.infer([data])[0]