File size: 14,829 Bytes
7af4264
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
import os
import random
import torch
import torch.utils.data
from tqdm import tqdm
from loguru import logger
import commons
from mel_processing import spectrogram_torch, mel_spectrogram_torch
from utils import load_filepaths_and_text
from utils import load_wav_to_torch_librosa as load_wav_to_torch
from text import cleaned_text_to_sequence, get_bert
import numpy as np

"""Multi speaker version"""


class TextAudioSpeakerLoader(torch.utils.data.Dataset):
    """
    1) loads audio, speaker_id, text pairs
    2) normalizes text and converts them to sequences of integers
    3) computes spectrograms from audio files.
    """

    def __init__(self, audiopaths_sid_text, hparams):
        self.audiopaths_sid_text = load_filepaths_and_text(audiopaths_sid_text)
        self.max_wav_value = hparams.max_wav_value
        self.sampling_rate = hparams.sampling_rate
        self.filter_length = hparams.filter_length
        self.hop_length = hparams.hop_length
        self.win_length = hparams.win_length
        self.sampling_rate = hparams.sampling_rate
        self.spk_map = hparams.spk2id
        self.hparams = hparams
        self.disable_bert = getattr(hparams, "disable_bert", False)

        self.use_mel_spec_posterior = getattr(
            hparams, "use_mel_posterior_encoder", False
        )
        if self.use_mel_spec_posterior:
            self.n_mel_channels = getattr(hparams, "n_mel_channels", 80)

        self.cleaned_text = getattr(hparams, "cleaned_text", False)

        self.add_blank = hparams.add_blank
        self.min_text_len = getattr(hparams, "min_text_len", 1)
        self.max_text_len = getattr(hparams, "max_text_len", 300)

        random.seed(1234)
        random.shuffle(self.audiopaths_sid_text)
        self._filter()


    def _filter(self):
        """
        Filter text & store spec lengths
        """
        # Store spectrogram lengths for Bucketing
        # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
        # spec_length = wav_length // hop_length

        audiopaths_sid_text_new = []
        lengths = []
        skipped = 0
        logger.info("Init dataset...")
        for item in tqdm(
            self.audiopaths_sid_text
        ):
            try:
                _id, spk, language, text, phones, tone, word2ph = item
            except:
                print(item)
                raise
            audiopath = f"{_id}"
            if self.min_text_len <= len(phones) and len(phones) <= self.max_text_len:
                phones = phones.split(" ")
                tone = [int(i) for i in tone.split(" ")]
                word2ph = [int(i) for i in word2ph.split(" ")]
                audiopaths_sid_text_new.append(
                    [audiopath, spk, language, text, phones, tone, word2ph]
                )
                lengths.append(os.path.getsize(audiopath) // (2 * self.hop_length))
            else:
                skipped += 1
        logger.info(f'min: {min(lengths)}; max: {max(lengths)}' )
        logger.info(
            "skipped: "
            + str(skipped)
            + ", total: "
            + str(len(self.audiopaths_sid_text))
        )
        self.audiopaths_sid_text = audiopaths_sid_text_new
        self.lengths = lengths

    def get_audio_text_speaker_pair(self, audiopath_sid_text):
        # separate filename, speaker_id and text
        audiopath, sid, language, text, phones, tone, word2ph = audiopath_sid_text

        bert, ja_bert, phones, tone, language = self.get_text(
            text, word2ph, phones, tone, language, audiopath
        )

        spec, wav = self.get_audio(audiopath)
        sid = int(getattr(self.spk_map, sid, '0'))
        sid = torch.LongTensor([sid])
        return (phones, spec, wav, sid, tone, language, bert, ja_bert)

    def get_audio(self, filename):
        audio_norm, sampling_rate = load_wav_to_torch(filename, self.sampling_rate)
        if sampling_rate != self.sampling_rate:
            raise ValueError(
                "{} {} SR doesn't match target {} SR".format(
                    filename, sampling_rate, self.sampling_rate
                )
            )
        # NOTE: normalize has been achieved by torchaudio
        # audio_norm = audio / self.max_wav_value
        audio_norm = audio_norm.unsqueeze(0)
        spec_filename = filename.replace(".wav", ".spec.pt")
        if self.use_mel_spec_posterior:
            spec_filename = spec_filename.replace(".spec.pt", ".mel.pt")
        try:
            spec = torch.load(spec_filename)
            assert False
        except:
            if self.use_mel_spec_posterior:
                spec = mel_spectrogram_torch(
                    audio_norm,
                    self.filter_length,
                    self.n_mel_channels,
                    self.sampling_rate,
                    self.hop_length,
                    self.win_length,
                    self.hparams.mel_fmin,
                    self.hparams.mel_fmax,
                    center=False,
                )
            else:
                spec = spectrogram_torch(
                    audio_norm,
                    self.filter_length,
                    self.sampling_rate,
                    self.hop_length,
                    self.win_length,
                    center=False,
                )
            spec = torch.squeeze(spec, 0)
            torch.save(spec, spec_filename)
        return spec, audio_norm

    def get_text(self, text, word2ph, phone, tone, language_str, wav_path):
        phone, tone, language = cleaned_text_to_sequence(phone, tone, language_str)
        if self.add_blank:
            phone = commons.intersperse(phone, 0)
            tone = commons.intersperse(tone, 0)
            language = commons.intersperse(language, 0)
            for i in range(len(word2ph)):
                word2ph[i] = word2ph[i] * 2
            word2ph[0] += 1
        bert_path = wav_path.replace(".wav", ".bert.pt")
        try:
            bert = torch.load(bert_path)
            assert bert.shape[-1] == len(phone)
        except Exception as e:
            print(e, wav_path, bert_path, bert.shape, len(phone))
            bert = get_bert(text, word2ph, language_str)
            torch.save(bert, bert_path)
            assert bert.shape[-1] == len(phone), phone

        if self.disable_bert:
            bert = torch.zeros(1024, len(phone))
            ja_bert = torch.zeros(768, len(phone))
        else:
            if language_str in ["ZH"]:
                bert = bert
                ja_bert = torch.zeros(768, len(phone))
            elif language_str in ["JP", "EN", "ZH_MIX_EN", "KR", 'SP', 'ES', 'FR', 'DE', 'RU']:
                ja_bert = bert
                bert = torch.zeros(1024, len(phone))
            else:
                raise
                bert = torch.zeros(1024, len(phone))
                ja_bert = torch.zeros(768, len(phone))
        assert bert.shape[-1] == len(phone)
        phone = torch.LongTensor(phone)
        tone = torch.LongTensor(tone)
        language = torch.LongTensor(language)
        return bert, ja_bert, phone, tone, language

    def get_sid(self, sid):
        sid = torch.LongTensor([int(sid)])
        return sid

    def __getitem__(self, index):
        return self.get_audio_text_speaker_pair(self.audiopaths_sid_text[index])

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


class TextAudioSpeakerCollate:
    """Zero-pads model inputs and targets"""

    def __init__(self, return_ids=False):
        self.return_ids = return_ids

    def __call__(self, batch):
        """Collate's training batch from normalized text, audio and speaker identities
        PARAMS
        ------
        batch: [text_normalized, spec_normalized, wav_normalized, sid]
        """
        # Right zero-pad all one-hot text sequences to max input length
        _, ids_sorted_decreasing = torch.sort(
            torch.LongTensor([x[1].size(1) for x in batch]), dim=0, descending=True
        )

        max_text_len = max([len(x[0]) for x in batch])
        max_spec_len = max([x[1].size(1) for x in batch])
        max_wav_len = max([x[2].size(1) for x in batch])

        text_lengths = torch.LongTensor(len(batch))
        spec_lengths = torch.LongTensor(len(batch))
        wav_lengths = torch.LongTensor(len(batch))
        sid = torch.LongTensor(len(batch))

        text_padded = torch.LongTensor(len(batch), max_text_len)
        tone_padded = torch.LongTensor(len(batch), max_text_len)
        language_padded = torch.LongTensor(len(batch), max_text_len)
        bert_padded = torch.FloatTensor(len(batch), 1024, max_text_len)
        ja_bert_padded = torch.FloatTensor(len(batch), 768, max_text_len)

        spec_padded = torch.FloatTensor(len(batch), batch[0][1].size(0), max_spec_len)
        wav_padded = torch.FloatTensor(len(batch), 1, max_wav_len)
        text_padded.zero_()
        tone_padded.zero_()
        language_padded.zero_()
        spec_padded.zero_()
        wav_padded.zero_()
        bert_padded.zero_()
        ja_bert_padded.zero_()
        for i in range(len(ids_sorted_decreasing)):
            row = batch[ids_sorted_decreasing[i]]

            text = row[0]
            text_padded[i, : text.size(0)] = text
            text_lengths[i] = text.size(0)

            spec = row[1]
            spec_padded[i, :, : spec.size(1)] = spec
            spec_lengths[i] = spec.size(1)

            wav = row[2]
            wav_padded[i, :, : wav.size(1)] = wav
            wav_lengths[i] = wav.size(1)

            sid[i] = row[3]

            tone = row[4]
            tone_padded[i, : tone.size(0)] = tone

            language = row[5]
            language_padded[i, : language.size(0)] = language

            bert = row[6]
            bert_padded[i, :, : bert.size(1)] = bert

            ja_bert = row[7]
            ja_bert_padded[i, :, : ja_bert.size(1)] = ja_bert

        return (
            text_padded,
            text_lengths,
            spec_padded,
            spec_lengths,
            wav_padded,
            wav_lengths,
            sid,
            tone_padded,
            language_padded,
            bert_padded,
            ja_bert_padded,
        )


class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
    """
    Maintain similar input lengths in a batch.
    Length groups are specified by boundaries.
    Ex) boundaries = [b1, b2, b3] -> any batch is included either {x | b1 < length(x) <=b2} or {x | b2 < length(x) <= b3}.

    It removes samples which are not included in the boundaries.
    Ex) boundaries = [b1, b2, b3] -> any x s.t. length(x) <= b1 or length(x) > b3 are discarded.
    """

    def __init__(
        self,
        dataset,
        batch_size,
        boundaries,
        num_replicas=None,
        rank=None,
        shuffle=True,
    ):
        super().__init__(dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle)
        self.lengths = dataset.lengths
        self.batch_size = batch_size
        self.boundaries = boundaries

        self.buckets, self.num_samples_per_bucket = self._create_buckets()
        self.total_size = sum(self.num_samples_per_bucket)
        self.num_samples = self.total_size // self.num_replicas
        print('buckets:', self.num_samples_per_bucket)

    def _create_buckets(self):
        buckets = [[] for _ in range(len(self.boundaries) - 1)]
        for i in range(len(self.lengths)):
            length = self.lengths[i]
            idx_bucket = self._bisect(length)
            if idx_bucket != -1:
                buckets[idx_bucket].append(i)

        try:
            for i in range(len(buckets) - 1, 0, -1):
                if len(buckets[i]) == 0:
                    buckets.pop(i)
                    self.boundaries.pop(i + 1)
            assert all(len(bucket) > 0 for bucket in buckets)
        # When one bucket is not traversed
        except Exception as e:
            print("Bucket warning ", e)
            for i in range(len(buckets) - 1, -1, -1):
                if len(buckets[i]) == 0:
                    buckets.pop(i)
                    self.boundaries.pop(i + 1)

        num_samples_per_bucket = []
        for i in range(len(buckets)):
            len_bucket = len(buckets[i])
            total_batch_size = self.num_replicas * self.batch_size
            rem = (
                total_batch_size - (len_bucket % total_batch_size)
            ) % total_batch_size
            num_samples_per_bucket.append(len_bucket + rem)
        return buckets, num_samples_per_bucket

    def __iter__(self):
        # deterministically shuffle based on epoch
        g = torch.Generator()
        g.manual_seed(self.epoch)

        indices = []
        if self.shuffle:
            for bucket in self.buckets:
                indices.append(torch.randperm(len(bucket), generator=g).tolist())
        else:
            for bucket in self.buckets:
                indices.append(list(range(len(bucket))))

        batches = []
        for i in range(len(self.buckets)):
            bucket = self.buckets[i]
            len_bucket = len(bucket)
            if len_bucket == 0:
                continue
            ids_bucket = indices[i]
            num_samples_bucket = self.num_samples_per_bucket[i]

            # add extra samples to make it evenly divisible
            rem = num_samples_bucket - len_bucket
            ids_bucket = (
                ids_bucket
                + ids_bucket * (rem // len_bucket)
                + ids_bucket[: (rem % len_bucket)]
            )

            # subsample
            ids_bucket = ids_bucket[self.rank :: self.num_replicas]

            # batching
            for j in range(len(ids_bucket) // self.batch_size):
                batch = [
                    bucket[idx]
                    for idx in ids_bucket[
                        j * self.batch_size : (j + 1) * self.batch_size
                    ]
                ]
                batches.append(batch)

        if self.shuffle:
            batch_ids = torch.randperm(len(batches), generator=g).tolist()
            batches = [batches[i] for i in batch_ids]
        self.batches = batches

        assert len(self.batches) * self.batch_size == self.num_samples
        return iter(self.batches)

    def _bisect(self, x, lo=0, hi=None):
        if hi is None:
            hi = len(self.boundaries) - 1

        if hi > lo:
            mid = (hi + lo) // 2
            if self.boundaries[mid] < x and x <= self.boundaries[mid + 1]:
                return mid
            elif x <= self.boundaries[mid]:
                return self._bisect(x, lo, mid)
            else:
                return self._bisect(x, mid + 1, hi)
        else:
            return -1

    def __len__(self):
        return self.num_samples // self.batch_size