File size: 10,389 Bytes
ee6e328
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import shutil
import unittest

import numpy as np
import torch
from torch.utils.data import DataLoader

from tests import get_tests_data_path, get_tests_output_path
from TTS.tts.configs.shared_configs import BaseDatasetConfig, BaseTTSConfig
from TTS.tts.datasets import TTSDataset, load_tts_samples
from TTS.tts.utils.text.tokenizer import TTSTokenizer
from TTS.utils.audio import AudioProcessor

# pylint: disable=unused-variable

OUTPATH = os.path.join(get_tests_output_path(), "loader_tests/")
os.makedirs(OUTPATH, exist_ok=True)

# create a dummy config for testing data loaders.
c = BaseTTSConfig(text_cleaner="english_cleaners", num_loader_workers=0, batch_size=2, use_noise_augment=False)
c.r = 5
c.data_path = os.path.join(get_tests_data_path(), "ljspeech/")
ok_ljspeech = os.path.exists(c.data_path)

dataset_config = BaseDatasetConfig(
    formatter="ljspeech_test",  # ljspeech_test to multi-speaker
    meta_file_train="metadata.csv",
    meta_file_val=None,
    path=c.data_path,
    language="en",
)

DATA_EXIST = True
if not os.path.exists(c.data_path):
    DATA_EXIST = False

print(" > Dynamic data loader test: {}".format(DATA_EXIST))


class TestTTSDataset(unittest.TestCase):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.max_loader_iter = 4
        self.ap = AudioProcessor(**c.audio)

    def _create_dataloader(self, batch_size, r, bgs, start_by_longest=False):
        # load dataset
        meta_data_train, meta_data_eval = load_tts_samples(dataset_config, eval_split=True, eval_split_size=0.2)
        items = meta_data_train + meta_data_eval

        tokenizer, _ = TTSTokenizer.init_from_config(c)
        dataset = TTSDataset(
            outputs_per_step=r,
            compute_linear_spec=True,
            return_wav=True,
            tokenizer=tokenizer,
            ap=self.ap,
            samples=items,
            batch_group_size=bgs,
            min_text_len=c.min_text_len,
            max_text_len=c.max_text_len,
            min_audio_len=c.min_audio_len,
            max_audio_len=c.max_audio_len,
            start_by_longest=start_by_longest,
        )
        dataloader = DataLoader(
            dataset,
            batch_size=batch_size,
            shuffle=False,
            collate_fn=dataset.collate_fn,
            drop_last=True,
            num_workers=c.num_loader_workers,
        )
        return dataloader, dataset

    def test_loader(self):
        if ok_ljspeech:
            dataloader, dataset = self._create_dataloader(1, 1, 0)

            for i, data in enumerate(dataloader):
                if i == self.max_loader_iter:
                    break
                text_input = data["token_id"]
                _ = data["token_id_lengths"]
                speaker_name = data["speaker_names"]
                linear_input = data["linear"]
                mel_input = data["mel"]
                mel_lengths = data["mel_lengths"]
                _ = data["stop_targets"]
                _ = data["item_idxs"]
                wavs = data["waveform"]

                neg_values = text_input[text_input < 0]
                check_count = len(neg_values)

                # check basic conditions
                self.assertEqual(check_count, 0)
                self.assertEqual(linear_input.shape[0], mel_input.shape[0], c.batch_size)
                self.assertEqual(linear_input.shape[2], self.ap.fft_size // 2 + 1)
                self.assertEqual(mel_input.shape[2], c.audio["num_mels"])
                self.assertEqual(wavs.shape[1], mel_input.shape[1] * c.audio.hop_length)
                self.assertIsInstance(speaker_name[0], str)

                # make sure that the computed mels and the waveform match and correctly computed
                mel_new = self.ap.melspectrogram(wavs[0].squeeze().numpy())
                # remove padding in mel-spectrogram
                mel_dataloader = mel_input[0].T.numpy()[:, : mel_lengths[0]]
                # guarantee that both mel-spectrograms have the same size and that we will remove waveform padding
                mel_new = mel_new[:, : mel_lengths[0]]
                ignore_seg = -(1 + c.audio.win_length // c.audio.hop_length)
                mel_diff = (mel_new[:, : mel_input.shape[1]] - mel_input[0].T.numpy())[:, 0:ignore_seg]
                self.assertLess(abs(mel_diff.sum()), 1e-5)

                # check normalization ranges
                if self.ap.symmetric_norm:
                    self.assertLessEqual(mel_input.max(), self.ap.max_norm)
                    self.assertGreaterEqual(
                        mel_input.min(), -self.ap.max_norm  # pylint: disable=invalid-unary-operand-type
                    )
                    self.assertLess(mel_input.min(), 0)
                else:
                    self.assertLessEqual(mel_input.max(), self.ap.max_norm)
                    self.assertGreaterEqual(mel_input.min(), 0)

    def test_batch_group_shuffle(self):
        if ok_ljspeech:
            dataloader, dataset = self._create_dataloader(2, c.r, 16)
            last_length = 0
            frames = dataset.samples
            for i, data in enumerate(dataloader):
                if i == self.max_loader_iter:
                    break
                mel_lengths = data["mel_lengths"]
                avg_length = mel_lengths.numpy().mean()
            dataloader.dataset.preprocess_samples()
            is_items_reordered = False
            for idx, item in enumerate(dataloader.dataset.samples):
                if item != frames[idx]:
                    is_items_reordered = True
                    break
            self.assertGreaterEqual(avg_length, last_length)
            self.assertTrue(is_items_reordered)

    def test_start_by_longest(self):
        """Test start_by_longest option.

        Ther first item of the fist batch must be longer than all the other items.
        """
        if ok_ljspeech:
            dataloader, _ = self._create_dataloader(2, c.r, 0, True)
            dataloader.dataset.preprocess_samples()
            for i, data in enumerate(dataloader):
                if i == self.max_loader_iter:
                    break
                mel_lengths = data["mel_lengths"]
                if i == 0:
                    max_len = mel_lengths[0]
                print(mel_lengths)
                self.assertTrue(all(max_len >= mel_lengths))

    def test_padding_and_spectrograms(self):
        def check_conditions(idx, linear_input, mel_input, stop_target, mel_lengths):
            self.assertNotEqual(linear_input[idx, -1].sum(), 0)  # check padding
            self.assertNotEqual(linear_input[idx, -2].sum(), 0)
            self.assertNotEqual(mel_input[idx, -1].sum(), 0)
            self.assertNotEqual(mel_input[idx, -2].sum(), 0)
            self.assertEqual(stop_target[idx, -1], 1)
            self.assertEqual(stop_target[idx, -2], 0)
            self.assertEqual(stop_target[idx].sum(), 1)
            self.assertEqual(len(mel_lengths.shape), 1)
            self.assertEqual(mel_lengths[idx], linear_input[idx].shape[0])
            self.assertEqual(mel_lengths[idx], mel_input[idx].shape[0])

        if ok_ljspeech:
            dataloader, _ = self._create_dataloader(1, 1, 0)

            for i, data in enumerate(dataloader):
                if i == self.max_loader_iter:
                    break
                linear_input = data["linear"]
                mel_input = data["mel"]
                mel_lengths = data["mel_lengths"]
                stop_target = data["stop_targets"]
                item_idx = data["item_idxs"]

                # check mel_spec consistency
                wav = np.asarray(self.ap.load_wav(item_idx[0]), dtype=np.float32)
                mel = self.ap.melspectrogram(wav).astype("float32")
                mel = torch.FloatTensor(mel).contiguous()
                mel_dl = mel_input[0]
                # NOTE: Below needs to check == 0 but due to an unknown reason
                # there is a slight difference between two matrices.
                # TODO: Check this assert cond more in detail.
                self.assertLess(abs(mel.T - mel_dl).max(), 1e-5)

                # check mel-spec correctness
                mel_spec = mel_input[0].cpu().numpy()
                wav = self.ap.inv_melspectrogram(mel_spec.T)
                self.ap.save_wav(wav, OUTPATH + "/mel_inv_dataloader.wav")
                shutil.copy(item_idx[0], OUTPATH + "/mel_target_dataloader.wav")

                # check linear-spec
                linear_spec = linear_input[0].cpu().numpy()
                wav = self.ap.inv_spectrogram(linear_spec.T)
                self.ap.save_wav(wav, OUTPATH + "/linear_inv_dataloader.wav")
                shutil.copy(item_idx[0], OUTPATH + "/linear_target_dataloader.wav")

                # check the outputs
                check_conditions(0, linear_input, mel_input, stop_target, mel_lengths)

            # Test for batch size 2
            dataloader, _ = self._create_dataloader(2, 1, 0)

            for i, data in enumerate(dataloader):
                if i == self.max_loader_iter:
                    break
                linear_input = data["linear"]
                mel_input = data["mel"]
                mel_lengths = data["mel_lengths"]
                stop_target = data["stop_targets"]
                item_idx = data["item_idxs"]

                # set id to the longest sequence in the batch
                if mel_lengths[0] > mel_lengths[1]:
                    idx = 0
                else:
                    idx = 1

                # check the longer item in the batch
                check_conditions(idx, linear_input, mel_input, stop_target, mel_lengths)

                # check the other item in the batch
                self.assertEqual(linear_input[1 - idx, -1].sum(), 0)
                self.assertEqual(mel_input[1 - idx, -1].sum(), 0)
                self.assertEqual(stop_target[1, mel_lengths[1] - 1], 1)
                self.assertEqual(stop_target[1, mel_lengths[1] :].sum(), stop_target.shape[1] - mel_lengths[1])
                self.assertEqual(len(mel_lengths.shape), 1)

                # check batch zero-frame conditions (zero-frame disabled)
                # assert (linear_input * stop_target.unsqueeze(2)).sum() == 0
                # assert (mel_input * stop_target.unsqueeze(2)).sum() == 0