justdata 1.0.0__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- justdata/__init__.py +3 -0
- justdata/acoustic/__init__.py +381 -0
- justdata/acoustic/adapters.py +329 -0
- justdata/acoustic/augment/__init__.py +119 -0
- justdata/acoustic/augment/batch.py +813 -0
- justdata/acoustic/augment/configs.py +204 -0
- justdata/acoustic/augment/patchout.py +331 -0
- justdata/acoustic/augment/spectrogram.py +806 -0
- justdata/acoustic/augment/waveform.py +879 -0
- justdata/acoustic/batching.py +27 -0
- justdata/acoustic/channel.py +53 -0
- justdata/acoustic/compat/__init__.py +95 -0
- justdata/acoustic/compat/ast.py +626 -0
- justdata/acoustic/compat/ced.py +152 -0
- justdata/acoustic/compat/efficientat.py +243 -0
- justdata/acoustic/compat/panns.py +115 -0
- justdata/acoustic/compat/passt.py +191 -0
- justdata/acoustic/configs.py +519 -0
- justdata/acoustic/corruptions/__init__.py +33 -0
- justdata/acoustic/corruptions/codec.py +92 -0
- justdata/acoustic/corruptions/datasets.py +155 -0
- justdata/acoustic/corruptions/device.py +39 -0
- justdata/acoustic/corruptions/dynamics.py +55 -0
- justdata/acoustic/corruptions/filters.py +105 -0
- justdata/acoustic/corruptions/noise.py +173 -0
- justdata/acoustic/corruptions/registry.py +230 -0
- justdata/acoustic/corruptions/reverb.py +49 -0
- justdata/acoustic/corruptions/timefreq.py +127 -0
- justdata/acoustic/datasets.py +52 -0
- justdata/acoustic/dcase2025.py +814 -0
- justdata/acoustic/decoding.py +73 -0
- justdata/acoustic/eval_views.py +54 -0
- justdata/acoustic/evals.py +71 -0
- justdata/acoustic/filters.py +18 -0
- justdata/acoustic/frontends/__init__.py +37 -0
- justdata/acoustic/frontends/ast.py +116 -0
- justdata/acoustic/frontends/kaldi.py +34 -0
- justdata/acoustic/frontends/log.py +94 -0
- justdata/acoustic/frontends/mel.py +173 -0
- justdata/acoustic/frontends/mfcc.py +26 -0
- justdata/acoustic/frontends/pcen.py +26 -0
- justdata/acoustic/frontends/stft.py +174 -0
- justdata/acoustic/jax.py +23 -0
- justdata/acoustic/labels.py +169 -0
- justdata/acoustic/layouts.py +86 -0
- justdata/acoustic/metadata.py +152 -0
- justdata/acoustic/normalization.py +196 -0
- justdata/acoustic/pipelines.py +112 -0
- justdata/acoustic/postprocessing.py +328 -0
- justdata/acoustic/preprocessing.py +175 -0
- justdata/acoustic/presets.py +257 -0
- justdata/acoustic/registry.py +381 -0
- justdata/acoustic/resampling.py +218 -0
- justdata/acoustic/schema.py +51 -0
- justdata/acoustic/segment.py +369 -0
- justdata/acoustic/sources.py +503 -0
- justdata/acoustic/stages.py +59 -0
- justdata/acoustic/stats.py +50 -0
- justdata/acoustic/tasks.py +375 -0
- justdata/audio/__init__.py +20 -0
- justdata/core/__init__.py +74 -0
- justdata/core/adapters.py +46 -0
- justdata/core/augmentations.py +34 -0
- justdata/core/filters.py +94 -0
- justdata/core/finalization.py +173 -0
- justdata/core/label_mixing.py +129 -0
- justdata/core/loader.py +384 -0
- justdata/core/metadata.py +250 -0
- justdata/core/presets.py +155 -0
- justdata/core/registry.py +229 -0
- justdata/core/sources.py +87 -0
- justdata/core/stats.py +227 -0
- justdata/corruptions.py +52 -0
- justdata/py.typed +0 -0
- justdata/vision/__init__.py +17 -0
- justdata/vision/augmentations/__init__.py +31 -0
- justdata/vision/augmentations/auto.py +480 -0
- justdata/vision/augmentations/color.py +163 -0
- justdata/vision/augmentations/composed.py +156 -0
- justdata/vision/augmentations/geometric.py +302 -0
- justdata/vision/augmentations/mixing.py +232 -0
- justdata/vision/augmentations/registry.py +142 -0
- justdata/vision/corruptions/__init__.py +16 -0
- justdata/vision/corruptions/blur.py +65 -0
- justdata/vision/corruptions/digital.py +26 -0
- justdata/vision/corruptions/noise.py +20 -0
- justdata/vision/corruptions/registry.py +113 -0
- justdata/vision/corruptions/weather.py +100 -0
- justdata/vision/datasets.py +34 -0
- justdata/vision/encodings/__init__.py +1 -0
- justdata/vision/encodings/heatmaps.py +196 -0
- justdata/vision/minic.py +69 -0
- justdata/vision/pipelines.py +62 -0
- justdata/vision/presets.py +509 -0
- justdata/vision/sources.py +1080 -0
- justdata/vision/stages.py +271 -0
- justdata/vision/tasks/__init__.py +1 -0
- justdata/vision/tasks/classification.py +321 -0
- justdata/vision/tasks/depth.py +25 -0
- justdata/vision/tasks/segmentation.py +194 -0
- justdata/vision/transforms.py +113 -0
- justdata/vision/utils.py +741 -0
- justdata-1.0.0.dist-info/METADATA +860 -0
- justdata-1.0.0.dist-info/RECORD +106 -0
- justdata-1.0.0.dist-info/WHEEL +4 -0
- justdata-1.0.0.dist-info/licenses/LICENSE +23 -0
|
@@ -0,0 +1,879 @@
|
|
|
1
|
+
from __future__ import annotations
|
|
2
|
+
|
|
3
|
+
import dataclasses
|
|
4
|
+
from collections.abc import Mapping, Sequence
|
|
5
|
+
from typing import Any
|
|
6
|
+
|
|
7
|
+
import tensorflow as tf
|
|
8
|
+
|
|
9
|
+
from justdata.acoustic.adapters import infer_duration
|
|
10
|
+
from justdata.acoustic.augment.configs import (
|
|
11
|
+
AdditiveNoiseConfig,
|
|
12
|
+
CodecSimulationConfig,
|
|
13
|
+
DynamicRangeCompressionConfig,
|
|
14
|
+
RIRConvolutionConfig,
|
|
15
|
+
RandomClippingConfig,
|
|
16
|
+
RandomGainConfig,
|
|
17
|
+
RandomTimeShiftConfig,
|
|
18
|
+
SpeedPerturbConfig,
|
|
19
|
+
)
|
|
20
|
+
from justdata.acoustic.registry import (
|
|
21
|
+
get_audio_waveform_augment,
|
|
22
|
+
register_audio_waveform_augment,
|
|
23
|
+
)
|
|
24
|
+
from justdata.acoustic.schema import DURATION, SAMPLE_RATE, WAVEFORM
|
|
25
|
+
|
|
26
|
+
|
|
27
|
+
_EPS = tf.constant(1e-8, dtype=tf.float32)
|
|
28
|
+
|
|
29
|
+
|
|
30
|
+
def _seed_tensor(seed: tf.Tensor | int | None) -> tf.Tensor:
|
|
31
|
+
if seed is None:
|
|
32
|
+
return tf.constant([0, 0], dtype=tf.int32)
|
|
33
|
+
|
|
34
|
+
seed = tf.cast(tf.convert_to_tensor(seed), tf.int32)
|
|
35
|
+
if seed.shape.rank == 0:
|
|
36
|
+
return tf.stack([seed, tf.constant(0, dtype=tf.int32)])
|
|
37
|
+
if seed.shape.rank == 1 and seed.shape[0] == 1:
|
|
38
|
+
return tf.stack([seed[0], tf.constant(0, dtype=tf.int32)])
|
|
39
|
+
return seed[:2]
|
|
40
|
+
|
|
41
|
+
|
|
42
|
+
def _as_waveform_tc(waveform: tf.Tensor) -> tuple[tf.Tensor, int | None]:
|
|
43
|
+
waveform = tf.convert_to_tensor(waveform)
|
|
44
|
+
rank = waveform.shape.rank
|
|
45
|
+
if rank == 1:
|
|
46
|
+
return tf.expand_dims(waveform, -1), rank
|
|
47
|
+
if rank != 2:
|
|
48
|
+
raise ValueError("Waveform augment input must have shape [T] or [T, C]")
|
|
49
|
+
return waveform, rank
|
|
50
|
+
|
|
51
|
+
|
|
52
|
+
def _restore_rank(waveform: tf.Tensor, rank: int | None) -> tf.Tensor:
|
|
53
|
+
if rank == 1:
|
|
54
|
+
return tf.squeeze(waveform, axis=-1)
|
|
55
|
+
return waveform
|
|
56
|
+
|
|
57
|
+
|
|
58
|
+
def _config_from(
|
|
59
|
+
config: Any,
|
|
60
|
+
config_cls: type,
|
|
61
|
+
overrides: Mapping[str, Any],
|
|
62
|
+
) -> Any:
|
|
63
|
+
data = {}
|
|
64
|
+
if config is not None:
|
|
65
|
+
if isinstance(config, config_cls):
|
|
66
|
+
data.update(dataclasses.asdict(config))
|
|
67
|
+
elif dataclasses.is_dataclass(config):
|
|
68
|
+
data.update(dataclasses.asdict(config))
|
|
69
|
+
elif isinstance(config, Mapping):
|
|
70
|
+
data.update(config)
|
|
71
|
+
else:
|
|
72
|
+
raise TypeError(f"config must be a {config_cls.__name__}, mapping, or None")
|
|
73
|
+
data.update({key: value for key, value in overrides.items() if value is not None})
|
|
74
|
+
return config_cls(**data)
|
|
75
|
+
|
|
76
|
+
|
|
77
|
+
def _enabled(is_training: bool, augment_eval: bool) -> bool:
|
|
78
|
+
return bool(is_training or augment_eval)
|
|
79
|
+
|
|
80
|
+
|
|
81
|
+
def _uniform_float(seed: tf.Tensor, minval: float, maxval: float) -> tf.Tensor:
|
|
82
|
+
if minval == maxval:
|
|
83
|
+
return tf.constant(minval, dtype=tf.float32)
|
|
84
|
+
return tf.random.stateless_uniform(
|
|
85
|
+
[],
|
|
86
|
+
seed=seed,
|
|
87
|
+
minval=tf.cast(minval, tf.float32),
|
|
88
|
+
maxval=tf.cast(maxval, tf.float32),
|
|
89
|
+
dtype=tf.float32,
|
|
90
|
+
)
|
|
91
|
+
|
|
92
|
+
|
|
93
|
+
def _uniform_int(seed: tf.Tensor, minval: tf.Tensor, maxval: tf.Tensor) -> tf.Tensor:
|
|
94
|
+
return tf.cond(
|
|
95
|
+
tf.equal(minval, maxval),
|
|
96
|
+
lambda: tf.cast(minval, tf.int32),
|
|
97
|
+
lambda: tf.random.stateless_uniform(
|
|
98
|
+
[],
|
|
99
|
+
seed=seed,
|
|
100
|
+
minval=tf.cast(minval, tf.int32),
|
|
101
|
+
maxval=tf.cast(maxval + 1, tf.int32),
|
|
102
|
+
dtype=tf.int32,
|
|
103
|
+
),
|
|
104
|
+
)
|
|
105
|
+
|
|
106
|
+
|
|
107
|
+
def _maybe_apply(
|
|
108
|
+
waveform: tf.Tensor,
|
|
109
|
+
prob: float,
|
|
110
|
+
seed: tf.Tensor,
|
|
111
|
+
apply_fn,
|
|
112
|
+
) -> tf.Tensor:
|
|
113
|
+
if prob <= 0.0:
|
|
114
|
+
return waveform
|
|
115
|
+
if prob >= 1.0:
|
|
116
|
+
return apply_fn()
|
|
117
|
+
should_apply = tf.random.stateless_uniform([], seed=seed) < tf.cast(
|
|
118
|
+
prob, tf.float32
|
|
119
|
+
)
|
|
120
|
+
return tf.cond(should_apply, apply_fn, lambda: waveform)
|
|
121
|
+
|
|
122
|
+
|
|
123
|
+
def _fit_length(audio: tf.Tensor, target_length: tf.Tensor) -> tf.Tensor:
|
|
124
|
+
target_length = tf.cast(tf.maximum(target_length, 1), tf.int32)
|
|
125
|
+
audio = audio[:target_length]
|
|
126
|
+
pad = tf.maximum(target_length - tf.shape(audio)[0], 0)
|
|
127
|
+
return tf.pad(audio, [[0, pad], [0, 0]])
|
|
128
|
+
|
|
129
|
+
|
|
130
|
+
def _sample_target_length(
|
|
131
|
+
sample_rate: tf.Tensor | int | None,
|
|
132
|
+
target_length: int | tf.Tensor | None,
|
|
133
|
+
crop_duration: float | None,
|
|
134
|
+
) -> tf.Tensor:
|
|
135
|
+
if target_length is not None:
|
|
136
|
+
return tf.cast(target_length, tf.int32)
|
|
137
|
+
if crop_duration is None:
|
|
138
|
+
raise ValueError("random_crop requires target_length or crop_duration")
|
|
139
|
+
if sample_rate is None:
|
|
140
|
+
raise ValueError("random_crop with crop_duration requires sample_rate")
|
|
141
|
+
target = tf.round(
|
|
142
|
+
tf.cast(crop_duration, tf.float32) * tf.cast(sample_rate, tf.float32)
|
|
143
|
+
)
|
|
144
|
+
return tf.cast(tf.maximum(target, 1.0), tf.int32)
|
|
145
|
+
|
|
146
|
+
|
|
147
|
+
def make_colored_noise(
|
|
148
|
+
shape: tf.Tensor | Sequence[int],
|
|
149
|
+
seed: tf.Tensor | int | None,
|
|
150
|
+
kind: str = "white",
|
|
151
|
+
) -> tf.Tensor:
|
|
152
|
+
"""Create unit-RMS white, pink, or brown noise with stateless randomness."""
|
|
153
|
+
seed = _seed_tensor(seed)
|
|
154
|
+
shape = tf.cast(tf.convert_to_tensor(shape), tf.int32)
|
|
155
|
+
noise = tf.random.stateless_normal(shape, seed=seed, dtype=tf.float32)
|
|
156
|
+
if kind == "white":
|
|
157
|
+
return _normalize_noise(noise)
|
|
158
|
+
|
|
159
|
+
audio, rank = _as_waveform_tc(noise)
|
|
160
|
+
if kind == "brown":
|
|
161
|
+
brown = tf.cumsum(audio, axis=0)
|
|
162
|
+
return _restore_rank(_normalize_noise(brown), rank)
|
|
163
|
+
if kind != "pink":
|
|
164
|
+
raise ValueError("noise kind must be one of 'white', 'pink', or 'brown'")
|
|
165
|
+
|
|
166
|
+
time = tf.shape(audio)[0]
|
|
167
|
+
channels_first = tf.transpose(audio, [1, 0])
|
|
168
|
+
spectrum = tf.signal.rfft(channels_first, fft_length=[time])
|
|
169
|
+
num_bins = tf.shape(spectrum)[-1]
|
|
170
|
+
freqs = tf.cast(tf.range(num_bins), tf.float32)
|
|
171
|
+
weights = tf.where(freqs > 0.0, tf.math.rsqrt(freqs), tf.zeros_like(freqs))
|
|
172
|
+
spectrum = spectrum * tf.cast(weights[tf.newaxis, :], spectrum.dtype)
|
|
173
|
+
pink = tf.signal.irfft(spectrum, fft_length=[time])
|
|
174
|
+
pink = tf.transpose(pink, [1, 0])
|
|
175
|
+
return _restore_rank(_normalize_noise(pink), rank)
|
|
176
|
+
|
|
177
|
+
|
|
178
|
+
def _normalize_noise(noise: tf.Tensor) -> tf.Tensor:
|
|
179
|
+
noise = tf.cast(noise, tf.float32)
|
|
180
|
+
axes = tf.range(tf.rank(noise))
|
|
181
|
+
noise = noise - tf.reduce_mean(noise, axis=axes, keepdims=True)
|
|
182
|
+
rms = tf.sqrt(tf.reduce_mean(tf.square(noise), axis=axes, keepdims=True))
|
|
183
|
+
return tf.math.divide_no_nan(noise, tf.maximum(rms, _EPS))
|
|
184
|
+
|
|
185
|
+
|
|
186
|
+
@register_audio_waveform_augment("random_gain")
|
|
187
|
+
def random_gain(
|
|
188
|
+
waveform: tf.Tensor,
|
|
189
|
+
sample_rate: tf.Tensor | int | None = None,
|
|
190
|
+
*,
|
|
191
|
+
seed: tf.Tensor | int | None = None,
|
|
192
|
+
config: RandomGainConfig | Mapping[str, Any] | None = None,
|
|
193
|
+
prob: float | None = None,
|
|
194
|
+
min_db: float | None = None,
|
|
195
|
+
max_db: float | None = None,
|
|
196
|
+
is_training: bool = True,
|
|
197
|
+
augment_eval: bool = False,
|
|
198
|
+
) -> tf.Tensor:
|
|
199
|
+
del sample_rate
|
|
200
|
+
if not _enabled(is_training, augment_eval):
|
|
201
|
+
return tf.convert_to_tensor(waveform)
|
|
202
|
+
|
|
203
|
+
cfg = _config_from(
|
|
204
|
+
config,
|
|
205
|
+
RandomGainConfig,
|
|
206
|
+
{"prob": prob, "min_db": min_db, "max_db": max_db},
|
|
207
|
+
)
|
|
208
|
+
seed = _seed_tensor(seed)
|
|
209
|
+
gate_seed, gain_seed = tf.unstack(tf.random.split(seed, 2))
|
|
210
|
+
input_dtype = tf.convert_to_tensor(waveform).dtype
|
|
211
|
+
audio = tf.cast(waveform, tf.float32)
|
|
212
|
+
|
|
213
|
+
def apply() -> tf.Tensor:
|
|
214
|
+
gain_db = _uniform_float(gain_seed, cfg.min_db, cfg.max_db)
|
|
215
|
+
scale = tf.pow(tf.constant(10.0, dtype=tf.float32), gain_db / 20.0)
|
|
216
|
+
return tf.cast(audio * scale, input_dtype)
|
|
217
|
+
|
|
218
|
+
return _maybe_apply(tf.convert_to_tensor(waveform), cfg.prob, gate_seed, apply)
|
|
219
|
+
|
|
220
|
+
|
|
221
|
+
@register_audio_waveform_augment("time_shift")
|
|
222
|
+
def time_shift(
|
|
223
|
+
waveform: tf.Tensor,
|
|
224
|
+
sample_rate: tf.Tensor | int | None = None,
|
|
225
|
+
*,
|
|
226
|
+
seed: tf.Tensor | int | None = None,
|
|
227
|
+
config: RandomTimeShiftConfig | Mapping[str, Any] | None = None,
|
|
228
|
+
prob: float | None = None,
|
|
229
|
+
max_shift_seconds: float | None = None,
|
|
230
|
+
mode: str | None = None,
|
|
231
|
+
shift_samples: int | tf.Tensor | None = None,
|
|
232
|
+
is_training: bool = True,
|
|
233
|
+
augment_eval: bool = False,
|
|
234
|
+
) -> tf.Tensor:
|
|
235
|
+
if not _enabled(is_training, augment_eval):
|
|
236
|
+
return tf.convert_to_tensor(waveform)
|
|
237
|
+
if sample_rate is None and shift_samples is None:
|
|
238
|
+
raise ValueError("time_shift requires sample_rate unless shift_samples is set")
|
|
239
|
+
|
|
240
|
+
cfg = _config_from(
|
|
241
|
+
config,
|
|
242
|
+
RandomTimeShiftConfig,
|
|
243
|
+
{"prob": prob, "max_shift_seconds": max_shift_seconds, "mode": mode},
|
|
244
|
+
)
|
|
245
|
+
seed = _seed_tensor(seed)
|
|
246
|
+
gate_seed, shift_seed = tf.unstack(tf.random.split(seed, 2))
|
|
247
|
+
audio, rank = _as_waveform_tc(waveform)
|
|
248
|
+
|
|
249
|
+
def apply() -> tf.Tensor:
|
|
250
|
+
if shift_samples is not None:
|
|
251
|
+
shift = tf.cast(shift_samples, tf.int32)
|
|
252
|
+
else:
|
|
253
|
+
max_shift = tf.cast(
|
|
254
|
+
tf.round(
|
|
255
|
+
tf.cast(cfg.max_shift_seconds, tf.float32)
|
|
256
|
+
* tf.cast(sample_rate, tf.float32)
|
|
257
|
+
),
|
|
258
|
+
tf.int32,
|
|
259
|
+
)
|
|
260
|
+
shift = _uniform_int(shift_seed, -max_shift, max_shift)
|
|
261
|
+
if cfg.mode == "roll":
|
|
262
|
+
shifted = tf.roll(audio, shift=shift, axis=0)
|
|
263
|
+
elif cfg.mode == "zero":
|
|
264
|
+
shifted = _zero_shift(audio, shift)
|
|
265
|
+
else:
|
|
266
|
+
raise ValueError(f"Unknown time shift mode: {cfg.mode!r}")
|
|
267
|
+
return _restore_rank(shifted, rank)
|
|
268
|
+
|
|
269
|
+
return _maybe_apply(tf.convert_to_tensor(waveform), cfg.prob, gate_seed, apply)
|
|
270
|
+
|
|
271
|
+
|
|
272
|
+
def _zero_shift(audio: tf.Tensor, shift: tf.Tensor) -> tf.Tensor:
|
|
273
|
+
time = tf.shape(audio)[0]
|
|
274
|
+
channels = tf.shape(audio)[1]
|
|
275
|
+
shift = tf.clip_by_value(shift, -time, time)
|
|
276
|
+
amount = tf.abs(shift)
|
|
277
|
+
zeros = tf.zeros([amount, channels], dtype=audio.dtype)
|
|
278
|
+
|
|
279
|
+
def positive() -> tf.Tensor:
|
|
280
|
+
return tf.concat([zeros, audio[: time - amount]], axis=0)
|
|
281
|
+
|
|
282
|
+
def negative() -> tf.Tensor:
|
|
283
|
+
return tf.concat([audio[amount:], zeros], axis=0)
|
|
284
|
+
|
|
285
|
+
return tf.cond(
|
|
286
|
+
shift > 0,
|
|
287
|
+
positive,
|
|
288
|
+
lambda: tf.cond(shift < 0, negative, lambda: audio),
|
|
289
|
+
)
|
|
290
|
+
|
|
291
|
+
|
|
292
|
+
@register_audio_waveform_augment("random_crop")
|
|
293
|
+
def random_crop(
|
|
294
|
+
waveform: tf.Tensor,
|
|
295
|
+
sample_rate: tf.Tensor | int | None = None,
|
|
296
|
+
*,
|
|
297
|
+
seed: tf.Tensor | int | None = None,
|
|
298
|
+
target_length: int | tf.Tensor | None = None,
|
|
299
|
+
crop_duration: float | None = None,
|
|
300
|
+
prob: float = 1.0,
|
|
301
|
+
is_training: bool = True,
|
|
302
|
+
augment_eval: bool = False,
|
|
303
|
+
) -> tf.Tensor:
|
|
304
|
+
if not _enabled(is_training, augment_eval):
|
|
305
|
+
return tf.convert_to_tensor(waveform)
|
|
306
|
+
|
|
307
|
+
seed = _seed_tensor(seed)
|
|
308
|
+
gate_seed, crop_seed = tf.unstack(tf.random.split(seed, 2))
|
|
309
|
+
audio, rank = _as_waveform_tc(waveform)
|
|
310
|
+
target = _sample_target_length(sample_rate, target_length, crop_duration)
|
|
311
|
+
|
|
312
|
+
def apply() -> tf.Tensor:
|
|
313
|
+
time = tf.shape(audio)[0]
|
|
314
|
+
max_start = tf.maximum(time - target, 0)
|
|
315
|
+
start = tf.cond(
|
|
316
|
+
max_start > 0,
|
|
317
|
+
lambda: tf.random.stateless_uniform(
|
|
318
|
+
[],
|
|
319
|
+
seed=crop_seed,
|
|
320
|
+
minval=0,
|
|
321
|
+
maxval=max_start + 1,
|
|
322
|
+
dtype=tf.int32,
|
|
323
|
+
),
|
|
324
|
+
lambda: tf.constant(0, dtype=tf.int32),
|
|
325
|
+
)
|
|
326
|
+
cropped = _fit_length(audio[start:], target)
|
|
327
|
+
return _restore_rank(cropped, rank)
|
|
328
|
+
|
|
329
|
+
return _maybe_apply(tf.convert_to_tensor(waveform), prob, gate_seed, apply)
|
|
330
|
+
|
|
331
|
+
|
|
332
|
+
@register_audio_waveform_augment("additive_noise")
|
|
333
|
+
def additive_noise(
|
|
334
|
+
waveform: tf.Tensor,
|
|
335
|
+
sample_rate: tf.Tensor | int | None = None,
|
|
336
|
+
*,
|
|
337
|
+
seed: tf.Tensor | int | None = None,
|
|
338
|
+
config: AdditiveNoiseConfig | Mapping[str, Any] | None = None,
|
|
339
|
+
prob: float | None = None,
|
|
340
|
+
snr_db_min: float | None = None,
|
|
341
|
+
snr_db_max: float | None = None,
|
|
342
|
+
noise_kind: str | None = None,
|
|
343
|
+
is_training: bool = True,
|
|
344
|
+
augment_eval: bool = False,
|
|
345
|
+
) -> tf.Tensor:
|
|
346
|
+
del sample_rate
|
|
347
|
+
if not _enabled(is_training, augment_eval):
|
|
348
|
+
return tf.convert_to_tensor(waveform)
|
|
349
|
+
|
|
350
|
+
cfg = _config_from(
|
|
351
|
+
config,
|
|
352
|
+
AdditiveNoiseConfig,
|
|
353
|
+
{
|
|
354
|
+
"prob": prob,
|
|
355
|
+
"snr_db_min": snr_db_min,
|
|
356
|
+
"snr_db_max": snr_db_max,
|
|
357
|
+
"noise_kind": noise_kind,
|
|
358
|
+
},
|
|
359
|
+
)
|
|
360
|
+
seed = _seed_tensor(seed)
|
|
361
|
+
gate_seed, snr_seed, noise_seed = tf.unstack(tf.random.split(seed, 3))
|
|
362
|
+
input_dtype = tf.convert_to_tensor(waveform).dtype
|
|
363
|
+
audio = tf.cast(waveform, tf.float32)
|
|
364
|
+
|
|
365
|
+
def apply() -> tf.Tensor:
|
|
366
|
+
noise = make_colored_noise(tf.shape(audio), noise_seed, cfg.noise_kind)
|
|
367
|
+
signal_power = tf.reduce_mean(tf.square(audio))
|
|
368
|
+
noise_power = tf.reduce_mean(tf.square(noise))
|
|
369
|
+
snr_db = _uniform_float(snr_seed, cfg.snr_db_min, cfg.snr_db_max)
|
|
370
|
+
target = tf.pow(tf.constant(10.0, dtype=tf.float32), snr_db / 10.0)
|
|
371
|
+
scale = tf.sqrt(tf.math.divide_no_nan(signal_power, noise_power * target))
|
|
372
|
+
return tf.cast(audio + noise * scale, input_dtype)
|
|
373
|
+
|
|
374
|
+
return _maybe_apply(tf.convert_to_tensor(waveform), cfg.prob, gate_seed, apply)
|
|
375
|
+
|
|
376
|
+
|
|
377
|
+
@register_audio_waveform_augment("colored_noise")
|
|
378
|
+
def colored_noise(
|
|
379
|
+
waveform: tf.Tensor,
|
|
380
|
+
sample_rate: tf.Tensor | int | None = None,
|
|
381
|
+
*,
|
|
382
|
+
seed: tf.Tensor | int | None = None,
|
|
383
|
+
config: AdditiveNoiseConfig | Mapping[str, Any] | None = None,
|
|
384
|
+
prob: float | None = None,
|
|
385
|
+
snr_db_min: float | None = None,
|
|
386
|
+
snr_db_max: float | None = None,
|
|
387
|
+
noise_kind: str = "pink",
|
|
388
|
+
is_training: bool = True,
|
|
389
|
+
augment_eval: bool = False,
|
|
390
|
+
) -> tf.Tensor:
|
|
391
|
+
return additive_noise(
|
|
392
|
+
waveform,
|
|
393
|
+
sample_rate,
|
|
394
|
+
seed=seed,
|
|
395
|
+
config=config,
|
|
396
|
+
prob=prob,
|
|
397
|
+
snr_db_min=snr_db_min,
|
|
398
|
+
snr_db_max=snr_db_max,
|
|
399
|
+
noise_kind=noise_kind,
|
|
400
|
+
is_training=is_training,
|
|
401
|
+
augment_eval=augment_eval,
|
|
402
|
+
)
|
|
403
|
+
|
|
404
|
+
|
|
405
|
+
@register_audio_waveform_augment("polarity_inversion")
|
|
406
|
+
def polarity_inversion(
|
|
407
|
+
waveform: tf.Tensor,
|
|
408
|
+
sample_rate: tf.Tensor | int | None = None,
|
|
409
|
+
*,
|
|
410
|
+
seed: tf.Tensor | int | None = None,
|
|
411
|
+
prob: float = 0.5,
|
|
412
|
+
is_training: bool = True,
|
|
413
|
+
augment_eval: bool = False,
|
|
414
|
+
) -> tf.Tensor:
|
|
415
|
+
del sample_rate
|
|
416
|
+
if not _enabled(is_training, augment_eval):
|
|
417
|
+
return tf.convert_to_tensor(waveform)
|
|
418
|
+
|
|
419
|
+
seed = _seed_tensor(seed)
|
|
420
|
+
audio = tf.convert_to_tensor(waveform)
|
|
421
|
+
return _maybe_apply(audio, prob, seed, lambda: -audio)
|
|
422
|
+
|
|
423
|
+
|
|
424
|
+
@register_audio_waveform_augment("random_clipping")
|
|
425
|
+
def random_clipping(
|
|
426
|
+
waveform: tf.Tensor,
|
|
427
|
+
sample_rate: tf.Tensor | int | None = None,
|
|
428
|
+
*,
|
|
429
|
+
seed: tf.Tensor | int | None = None,
|
|
430
|
+
config: RandomClippingConfig | Mapping[str, Any] | None = None,
|
|
431
|
+
prob: float | None = None,
|
|
432
|
+
min_threshold: float | None = None,
|
|
433
|
+
max_threshold: float | None = None,
|
|
434
|
+
is_training: bool = True,
|
|
435
|
+
augment_eval: bool = False,
|
|
436
|
+
) -> tf.Tensor:
|
|
437
|
+
del sample_rate
|
|
438
|
+
if not _enabled(is_training, augment_eval):
|
|
439
|
+
return tf.convert_to_tensor(waveform)
|
|
440
|
+
|
|
441
|
+
cfg = _config_from(
|
|
442
|
+
config,
|
|
443
|
+
RandomClippingConfig,
|
|
444
|
+
{
|
|
445
|
+
"prob": prob,
|
|
446
|
+
"min_threshold": min_threshold,
|
|
447
|
+
"max_threshold": max_threshold,
|
|
448
|
+
},
|
|
449
|
+
)
|
|
450
|
+
seed = _seed_tensor(seed)
|
|
451
|
+
gate_seed, threshold_seed = tf.unstack(tf.random.split(seed, 2))
|
|
452
|
+
audio = tf.convert_to_tensor(waveform)
|
|
453
|
+
|
|
454
|
+
def apply() -> tf.Tensor:
|
|
455
|
+
threshold = tf.cast(
|
|
456
|
+
_uniform_float(threshold_seed, cfg.min_threshold, cfg.max_threshold),
|
|
457
|
+
audio.dtype,
|
|
458
|
+
)
|
|
459
|
+
return tf.clip_by_value(audio, -threshold, threshold)
|
|
460
|
+
|
|
461
|
+
return _maybe_apply(audio, cfg.prob, gate_seed, apply)
|
|
462
|
+
|
|
463
|
+
|
|
464
|
+
@register_audio_waveform_augment("dynamic_range_compression")
|
|
465
|
+
def dynamic_range_compression(
|
|
466
|
+
waveform: tf.Tensor,
|
|
467
|
+
sample_rate: tf.Tensor | int | None = None,
|
|
468
|
+
*,
|
|
469
|
+
seed: tf.Tensor | int | None = None,
|
|
470
|
+
config: DynamicRangeCompressionConfig | Mapping[str, Any] | None = None,
|
|
471
|
+
prob: float | None = None,
|
|
472
|
+
threshold_db_min: float | None = None,
|
|
473
|
+
threshold_db_max: float | None = None,
|
|
474
|
+
ratio_min: float | None = None,
|
|
475
|
+
ratio_max: float | None = None,
|
|
476
|
+
eps: float = 1e-8,
|
|
477
|
+
is_training: bool = True,
|
|
478
|
+
augment_eval: bool = False,
|
|
479
|
+
) -> tf.Tensor:
|
|
480
|
+
del sample_rate
|
|
481
|
+
if not _enabled(is_training, augment_eval):
|
|
482
|
+
return tf.convert_to_tensor(waveform)
|
|
483
|
+
|
|
484
|
+
cfg = _config_from(
|
|
485
|
+
config,
|
|
486
|
+
DynamicRangeCompressionConfig,
|
|
487
|
+
{
|
|
488
|
+
"prob": prob,
|
|
489
|
+
"threshold_db_min": threshold_db_min,
|
|
490
|
+
"threshold_db_max": threshold_db_max,
|
|
491
|
+
"ratio_min": ratio_min,
|
|
492
|
+
"ratio_max": ratio_max,
|
|
493
|
+
},
|
|
494
|
+
)
|
|
495
|
+
seed = _seed_tensor(seed)
|
|
496
|
+
gate_seed, threshold_seed, ratio_seed = tf.unstack(tf.random.split(seed, 3))
|
|
497
|
+
input_dtype = tf.convert_to_tensor(waveform).dtype
|
|
498
|
+
audio = tf.cast(waveform, tf.float32)
|
|
499
|
+
|
|
500
|
+
def apply() -> tf.Tensor:
|
|
501
|
+
threshold_db = _uniform_float(
|
|
502
|
+
threshold_seed,
|
|
503
|
+
cfg.threshold_db_min,
|
|
504
|
+
cfg.threshold_db_max,
|
|
505
|
+
)
|
|
506
|
+
ratio = _uniform_float(ratio_seed, cfg.ratio_min, cfg.ratio_max)
|
|
507
|
+
magnitude = tf.abs(audio)
|
|
508
|
+
level_db = (
|
|
509
|
+
20.0 * tf.math.log(magnitude + tf.cast(eps, tf.float32)) / tf.math.log(10.0)
|
|
510
|
+
)
|
|
511
|
+
over = level_db - threshold_db
|
|
512
|
+
compressed_over = over / ratio
|
|
513
|
+
gain_db = tf.where(over > 0.0, compressed_over - over, tf.zeros_like(over))
|
|
514
|
+
gain = tf.pow(tf.constant(10.0, dtype=tf.float32), gain_db / 20.0)
|
|
515
|
+
return tf.cast(tf.sign(audio) * magnitude * gain, input_dtype)
|
|
516
|
+
|
|
517
|
+
return _maybe_apply(tf.convert_to_tensor(waveform), cfg.prob, gate_seed, apply)
|
|
518
|
+
|
|
519
|
+
|
|
520
|
+
@register_audio_waveform_augment("rir_convolution")
|
|
521
|
+
def rir_convolution(
|
|
522
|
+
waveform: tf.Tensor,
|
|
523
|
+
sample_rate: tf.Tensor | int | None = None,
|
|
524
|
+
*,
|
|
525
|
+
seed: tf.Tensor | int | None = None,
|
|
526
|
+
config: RIRConvolutionConfig | Mapping[str, Any] | None = None,
|
|
527
|
+
prob: float | None = None,
|
|
528
|
+
rir: tf.Tensor | Sequence[float] | None = None,
|
|
529
|
+
rir_dataset: str | None = None,
|
|
530
|
+
normalize_ir: bool | None = None,
|
|
531
|
+
compensate_delay: bool | None = None,
|
|
532
|
+
is_training: bool = True,
|
|
533
|
+
augment_eval: bool = False,
|
|
534
|
+
) -> tf.Tensor:
|
|
535
|
+
del sample_rate
|
|
536
|
+
if not _enabled(is_training, augment_eval):
|
|
537
|
+
return tf.convert_to_tensor(waveform)
|
|
538
|
+
|
|
539
|
+
cfg = _config_from(
|
|
540
|
+
config,
|
|
541
|
+
RIRConvolutionConfig,
|
|
542
|
+
{
|
|
543
|
+
"prob": prob,
|
|
544
|
+
"rir_dataset": rir_dataset,
|
|
545
|
+
"normalize_ir": normalize_ir,
|
|
546
|
+
"compensate_delay": compensate_delay,
|
|
547
|
+
},
|
|
548
|
+
)
|
|
549
|
+
if cfg.rir_dataset is not None and rir is None:
|
|
550
|
+
raise ValueError(
|
|
551
|
+
"rir_dataset loading is not implemented; pass an explicit rir tensor"
|
|
552
|
+
)
|
|
553
|
+
|
|
554
|
+
seed = _seed_tensor(seed)
|
|
555
|
+
gate_seed, rir_seed = tf.unstack(tf.random.split(seed, 2))
|
|
556
|
+
input_dtype = tf.convert_to_tensor(waveform).dtype
|
|
557
|
+
audio, rank = _as_waveform_tc(tf.cast(waveform, tf.float32))
|
|
558
|
+
ir = _prepare_rir(rir, rir_seed, cfg.normalize_ir)
|
|
559
|
+
|
|
560
|
+
def apply() -> tf.Tensor:
|
|
561
|
+
convolved = _convolve_channels(audio, ir, cfg.compensate_delay)
|
|
562
|
+
return tf.cast(_restore_rank(convolved, rank), input_dtype)
|
|
563
|
+
|
|
564
|
+
return _maybe_apply(tf.convert_to_tensor(waveform), cfg.prob, gate_seed, apply)
|
|
565
|
+
|
|
566
|
+
|
|
567
|
+
def _prepare_rir(
|
|
568
|
+
rir: tf.Tensor | Sequence[float] | None,
|
|
569
|
+
seed: tf.Tensor,
|
|
570
|
+
normalize: bool,
|
|
571
|
+
) -> tf.Tensor:
|
|
572
|
+
if rir is None:
|
|
573
|
+
tail = tf.random.stateless_normal(
|
|
574
|
+
[63], seed=seed, stddev=0.05, dtype=tf.float32
|
|
575
|
+
)
|
|
576
|
+
decay = tf.exp(-tf.linspace(0.0, 4.0, 63))
|
|
577
|
+
rir = tf.concat([tf.ones([1], dtype=tf.float32), tail * decay], axis=0)
|
|
578
|
+
ir = tf.reshape(tf.cast(tf.convert_to_tensor(rir), tf.float32), [-1])
|
|
579
|
+
ir = tf.cond(
|
|
580
|
+
tf.shape(ir)[0] > 0, lambda: ir, lambda: tf.ones([1], dtype=tf.float32)
|
|
581
|
+
)
|
|
582
|
+
if normalize:
|
|
583
|
+
energy = tf.sqrt(tf.reduce_sum(tf.square(ir)))
|
|
584
|
+
ir = tf.math.divide_no_nan(ir, tf.maximum(energy, _EPS))
|
|
585
|
+
return ir
|
|
586
|
+
|
|
587
|
+
|
|
588
|
+
def _convolve_channels(
|
|
589
|
+
audio: tf.Tensor, ir: tf.Tensor, compensate_delay: bool
|
|
590
|
+
) -> tf.Tensor:
|
|
591
|
+
time = tf.shape(audio)[0]
|
|
592
|
+
kernel = tf.reverse(ir, axis=[0])[:, tf.newaxis, tf.newaxis]
|
|
593
|
+
kernel_length = tf.shape(kernel)[0]
|
|
594
|
+
start = (
|
|
595
|
+
tf.argmax(tf.abs(ir), output_type=tf.int32)
|
|
596
|
+
if compensate_delay
|
|
597
|
+
else (kernel_length - 1) // 2
|
|
598
|
+
)
|
|
599
|
+
|
|
600
|
+
def convolve_channel(channel: tf.Tensor) -> tf.Tensor:
|
|
601
|
+
signal = channel[tf.newaxis, :, tf.newaxis]
|
|
602
|
+
padded = tf.pad(
|
|
603
|
+
signal, [[0, 0], [kernel_length - 1, kernel_length - 1], [0, 0]]
|
|
604
|
+
)
|
|
605
|
+
full = tf.nn.conv1d(padded, kernel, stride=1, padding="VALID")[0, :, 0]
|
|
606
|
+
return _fit_length(full[start:, tf.newaxis], time)[:, 0]
|
|
607
|
+
|
|
608
|
+
channels_first = tf.transpose(audio, [1, 0])
|
|
609
|
+
convolved = tf.map_fn(
|
|
610
|
+
convolve_channel,
|
|
611
|
+
channels_first,
|
|
612
|
+
fn_output_signature=tf.float32,
|
|
613
|
+
)
|
|
614
|
+
return tf.transpose(convolved, [1, 0])
|
|
615
|
+
|
|
616
|
+
|
|
617
|
+
@register_audio_waveform_augment("speed_perturb")
|
|
618
|
+
def speed_perturb(
|
|
619
|
+
waveform: tf.Tensor,
|
|
620
|
+
sample_rate: tf.Tensor | int | None = None,
|
|
621
|
+
*,
|
|
622
|
+
seed: tf.Tensor | int | None = None,
|
|
623
|
+
config: SpeedPerturbConfig | Mapping[str, Any] | None = None,
|
|
624
|
+
prob: float | None = None,
|
|
625
|
+
rates: Sequence[float] | None = None,
|
|
626
|
+
is_training: bool = True,
|
|
627
|
+
augment_eval: bool = False,
|
|
628
|
+
) -> tf.Tensor:
|
|
629
|
+
del sample_rate
|
|
630
|
+
if not _enabled(is_training, augment_eval):
|
|
631
|
+
return tf.convert_to_tensor(waveform)
|
|
632
|
+
|
|
633
|
+
cfg = _config_from(
|
|
634
|
+
config,
|
|
635
|
+
SpeedPerturbConfig,
|
|
636
|
+
{"prob": prob, "rates": tuple(rates) if rates is not None else None},
|
|
637
|
+
)
|
|
638
|
+
seed = _seed_tensor(seed)
|
|
639
|
+
gate_seed, rate_seed = tf.unstack(tf.random.split(seed, 2))
|
|
640
|
+
input_dtype = tf.convert_to_tensor(waveform).dtype
|
|
641
|
+
audio, rank = _as_waveform_tc(tf.cast(waveform, tf.float32))
|
|
642
|
+
|
|
643
|
+
def apply() -> tf.Tensor:
|
|
644
|
+
rates_tensor = tf.constant(cfg.rates, dtype=tf.float32)
|
|
645
|
+
if len(cfg.rates) == 1:
|
|
646
|
+
rate = rates_tensor[0]
|
|
647
|
+
else:
|
|
648
|
+
index = tf.random.stateless_uniform(
|
|
649
|
+
[],
|
|
650
|
+
seed=rate_seed,
|
|
651
|
+
minval=0,
|
|
652
|
+
maxval=len(cfg.rates),
|
|
653
|
+
dtype=tf.int32,
|
|
654
|
+
)
|
|
655
|
+
rate = rates_tensor[index]
|
|
656
|
+
time = tf.shape(audio)[0]
|
|
657
|
+
new_time = tf.cast(
|
|
658
|
+
tf.maximum(tf.round(tf.cast(time, tf.float32) / rate), 1.0),
|
|
659
|
+
tf.int32,
|
|
660
|
+
)
|
|
661
|
+
resized = _resize_time(audio, new_time)
|
|
662
|
+
return tf.cast(_restore_rank(_fit_length(resized, time), rank), input_dtype)
|
|
663
|
+
|
|
664
|
+
return _maybe_apply(tf.convert_to_tensor(waveform), cfg.prob, gate_seed, apply)
|
|
665
|
+
|
|
666
|
+
|
|
667
|
+
def _resize_time(audio: tf.Tensor, new_time: tf.Tensor) -> tf.Tensor:
|
|
668
|
+
resized = tf.image.resize(
|
|
669
|
+
audio[tf.newaxis, :, tf.newaxis, :],
|
|
670
|
+
[new_time, 1],
|
|
671
|
+
method="bilinear",
|
|
672
|
+
antialias=True,
|
|
673
|
+
)
|
|
674
|
+
return resized[0, :, 0, :]
|
|
675
|
+
|
|
676
|
+
|
|
677
|
+
@register_audio_waveform_augment("codec_simulation")
|
|
678
|
+
def codec_simulation(
|
|
679
|
+
waveform: tf.Tensor,
|
|
680
|
+
sample_rate: tf.Tensor | int | None = None,
|
|
681
|
+
*,
|
|
682
|
+
seed: tf.Tensor | int | None = None,
|
|
683
|
+
config: CodecSimulationConfig | Mapping[str, Any] | None = None,
|
|
684
|
+
prob: float | None = None,
|
|
685
|
+
codec: str | None = None,
|
|
686
|
+
bitrate: int | None = None,
|
|
687
|
+
is_training: bool = True,
|
|
688
|
+
augment_eval: bool = False,
|
|
689
|
+
) -> tf.Tensor:
|
|
690
|
+
del sample_rate
|
|
691
|
+
if not _enabled(is_training, augment_eval):
|
|
692
|
+
return tf.convert_to_tensor(waveform)
|
|
693
|
+
|
|
694
|
+
cfg = _config_from(
|
|
695
|
+
config,
|
|
696
|
+
CodecSimulationConfig,
|
|
697
|
+
{"prob": prob, "codec": codec, "bitrate": bitrate},
|
|
698
|
+
)
|
|
699
|
+
gate_seed = _seed_tensor(seed)
|
|
700
|
+
input_dtype = tf.convert_to_tensor(waveform).dtype
|
|
701
|
+
audio, rank = _as_waveform_tc(tf.cast(waveform, tf.float32))
|
|
702
|
+
|
|
703
|
+
def apply() -> tf.Tensor:
|
|
704
|
+
clipped = tf.clip_by_value(audio, -1.0, 1.0)
|
|
705
|
+
if cfg.codec == "mulaw":
|
|
706
|
+
coded = _mulaw_roundtrip(clipped)
|
|
707
|
+
elif cfg.codec == "alaw":
|
|
708
|
+
coded = _alaw_roundtrip(clipped)
|
|
709
|
+
elif cfg.codec == "mp3_proxy":
|
|
710
|
+
coded = _mp3_proxy(clipped, cfg.bitrate)
|
|
711
|
+
else:
|
|
712
|
+
raise ValueError(f"Unknown codec simulation: {cfg.codec!r}")
|
|
713
|
+
return tf.cast(
|
|
714
|
+
_restore_rank(tf.clip_by_value(coded, -1.0, 1.0), rank), input_dtype
|
|
715
|
+
)
|
|
716
|
+
|
|
717
|
+
return _maybe_apply(tf.convert_to_tensor(waveform), cfg.prob, gate_seed, apply)
|
|
718
|
+
|
|
719
|
+
|
|
720
|
+
def _quantize_unit(x: tf.Tensor, levels: int) -> tf.Tensor:
|
|
721
|
+
levels_f = tf.cast(levels - 1, tf.float32)
|
|
722
|
+
x = tf.clip_by_value(x, -1.0, 1.0)
|
|
723
|
+
return tf.round((x + 1.0) * 0.5 * levels_f) / levels_f * 2.0 - 1.0
|
|
724
|
+
|
|
725
|
+
|
|
726
|
+
def _mulaw_roundtrip(audio: tf.Tensor) -> tf.Tensor:
|
|
727
|
+
mu = tf.constant(255.0, dtype=tf.float32)
|
|
728
|
+
log_mu = tf.math.log1p(mu)
|
|
729
|
+
encoded = tf.sign(audio) * tf.math.log1p(mu * tf.abs(audio)) / log_mu
|
|
730
|
+
encoded = _quantize_unit(encoded, 256)
|
|
731
|
+
return tf.sign(encoded) * tf.math.expm1(tf.abs(encoded) * log_mu) / mu
|
|
732
|
+
|
|
733
|
+
|
|
734
|
+
def _alaw_roundtrip(audio: tf.Tensor) -> tf.Tensor:
|
|
735
|
+
a = tf.constant(87.6, dtype=tf.float32)
|
|
736
|
+
denom = 1.0 + tf.math.log(a)
|
|
737
|
+
abs_audio = tf.abs(audio)
|
|
738
|
+
encoded_mag = tf.where(
|
|
739
|
+
abs_audio < (1.0 / a),
|
|
740
|
+
a * abs_audio / denom,
|
|
741
|
+
(1.0 + tf.math.log(a * tf.maximum(abs_audio, _EPS))) / denom,
|
|
742
|
+
)
|
|
743
|
+
encoded = _quantize_unit(tf.sign(audio) * encoded_mag, 256)
|
|
744
|
+
abs_encoded = tf.abs(encoded)
|
|
745
|
+
decoded_mag = tf.where(
|
|
746
|
+
abs_encoded < (1.0 / denom),
|
|
747
|
+
abs_encoded * denom / a,
|
|
748
|
+
tf.exp(abs_encoded * denom - 1.0) / a,
|
|
749
|
+
)
|
|
750
|
+
return tf.sign(encoded) * decoded_mag
|
|
751
|
+
|
|
752
|
+
|
|
753
|
+
def _mp3_proxy(audio: tf.Tensor, bitrate: int | None) -> tf.Tensor:
|
|
754
|
+
del bitrate
|
|
755
|
+
kernel = tf.constant([0.2, 0.6, 0.2], dtype=tf.float32)[:, tf.newaxis, tf.newaxis]
|
|
756
|
+
|
|
757
|
+
def smooth_channel(channel: tf.Tensor) -> tf.Tensor:
|
|
758
|
+
signal = channel[tf.newaxis, :, tf.newaxis]
|
|
759
|
+
return tf.nn.conv1d(signal, kernel, stride=1, padding="SAME")[0, :, 0]
|
|
760
|
+
|
|
761
|
+
smoothed = tf.map_fn(
|
|
762
|
+
smooth_channel,
|
|
763
|
+
tf.transpose(audio, [1, 0]),
|
|
764
|
+
fn_output_signature=tf.float32,
|
|
765
|
+
)
|
|
766
|
+
smoothed = tf.transpose(smoothed, [1, 0])
|
|
767
|
+
return _quantize_unit(smoothed, 128)
|
|
768
|
+
|
|
769
|
+
|
|
770
|
+
def normalize_waveform_augment_specs(
|
|
771
|
+
augmentations: Mapping[str, Any] | Sequence[Any] | str | None,
|
|
772
|
+
) -> list[dict[str, Any]]:
|
|
773
|
+
if augmentations is None:
|
|
774
|
+
return []
|
|
775
|
+
if isinstance(augmentations, str):
|
|
776
|
+
return [{"name": augmentations}]
|
|
777
|
+
if isinstance(augmentations, Mapping):
|
|
778
|
+
if "name" in augmentations:
|
|
779
|
+
spec = dict(augmentations)
|
|
780
|
+
spec["name"] = str(spec["name"])
|
|
781
|
+
return [spec]
|
|
782
|
+
|
|
783
|
+
specs = []
|
|
784
|
+
for name, value in augmentations.items():
|
|
785
|
+
if value is None or value is False:
|
|
786
|
+
continue
|
|
787
|
+
if value is True:
|
|
788
|
+
specs.append({"name": str(name)})
|
|
789
|
+
elif isinstance(value, Mapping):
|
|
790
|
+
specs.append({"name": str(name), **dict(value)})
|
|
791
|
+
else:
|
|
792
|
+
specs.append({"name": str(name), "config": value})
|
|
793
|
+
return specs
|
|
794
|
+
|
|
795
|
+
specs = []
|
|
796
|
+
for value in augmentations:
|
|
797
|
+
specs.extend(normalize_waveform_augment_specs(value))
|
|
798
|
+
return specs
|
|
799
|
+
|
|
800
|
+
|
|
801
|
+
def apply_waveform_augmentations(
|
|
802
|
+
sample: dict,
|
|
803
|
+
augmentations: Mapping[str, Any] | Sequence[Any] | str | None,
|
|
804
|
+
*,
|
|
805
|
+
seed: tf.Tensor | int | None = None,
|
|
806
|
+
is_training: bool = True,
|
|
807
|
+
augment_eval: bool = False,
|
|
808
|
+
audio_key: str = WAVEFORM,
|
|
809
|
+
) -> dict:
|
|
810
|
+
specs = normalize_waveform_augment_specs(augmentations)
|
|
811
|
+
if not specs or not _enabled(is_training, augment_eval):
|
|
812
|
+
return sample
|
|
813
|
+
|
|
814
|
+
seed = _seed_tensor(seed)
|
|
815
|
+
seeds = tf.random.split(seed, len(specs))
|
|
816
|
+
waveform = sample[audio_key]
|
|
817
|
+
sample_rate = sample.get(SAMPLE_RATE)
|
|
818
|
+
|
|
819
|
+
for spec, spec_seed in zip(specs, tf.unstack(seeds)):
|
|
820
|
+
kwargs = dict(spec)
|
|
821
|
+
name = kwargs.pop("name")
|
|
822
|
+
fn = get_audio_waveform_augment(name)
|
|
823
|
+
waveform = fn(
|
|
824
|
+
waveform,
|
|
825
|
+
sample_rate,
|
|
826
|
+
seed=spec_seed,
|
|
827
|
+
is_training=is_training,
|
|
828
|
+
augment_eval=augment_eval,
|
|
829
|
+
**kwargs,
|
|
830
|
+
)
|
|
831
|
+
|
|
832
|
+
result = dict(sample)
|
|
833
|
+
result[audio_key] = waveform
|
|
834
|
+
if DURATION in result and sample_rate is not None:
|
|
835
|
+
result[DURATION] = infer_duration(
|
|
836
|
+
tf.shape(tf.convert_to_tensor(waveform))[0], sample_rate
|
|
837
|
+
)
|
|
838
|
+
return result
|
|
839
|
+
|
|
840
|
+
|
|
841
|
+
def make_waveform_augmentation_stage(
|
|
842
|
+
augmentations: Mapping[str, Any] | Sequence[Any] | str | None,
|
|
843
|
+
*,
|
|
844
|
+
is_training: bool = True,
|
|
845
|
+
augment_eval: bool = False,
|
|
846
|
+
audio_key: str = WAVEFORM,
|
|
847
|
+
):
|
|
848
|
+
specs = normalize_waveform_augment_specs(augmentations)
|
|
849
|
+
|
|
850
|
+
def stage(sample: dict, seed: tf.Tensor | int | None = None) -> dict:
|
|
851
|
+
return apply_waveform_augmentations(
|
|
852
|
+
sample,
|
|
853
|
+
specs,
|
|
854
|
+
seed=seed,
|
|
855
|
+
is_training=is_training,
|
|
856
|
+
augment_eval=augment_eval,
|
|
857
|
+
audio_key=audio_key,
|
|
858
|
+
)
|
|
859
|
+
|
|
860
|
+
return stage
|
|
861
|
+
|
|
862
|
+
|
|
863
|
+
__all__ = [
|
|
864
|
+
"additive_noise",
|
|
865
|
+
"apply_waveform_augmentations",
|
|
866
|
+
"codec_simulation",
|
|
867
|
+
"colored_noise",
|
|
868
|
+
"dynamic_range_compression",
|
|
869
|
+
"make_colored_noise",
|
|
870
|
+
"make_waveform_augmentation_stage",
|
|
871
|
+
"normalize_waveform_augment_specs",
|
|
872
|
+
"polarity_inversion",
|
|
873
|
+
"random_clipping",
|
|
874
|
+
"random_crop",
|
|
875
|
+
"random_gain",
|
|
876
|
+
"rir_convolution",
|
|
877
|
+
"speed_perturb",
|
|
878
|
+
"time_shift",
|
|
879
|
+
]
|