File size: 15,689 Bytes
f4fa8f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from math import acos, sin
from typing import Iterable, Tuple, Union, List

import torch
import numpy as np
from PIL import Image
from tqdm.auto import tqdm
from librosa.beat import beat_track
from diffusers import (DiffusionPipeline, UNet2DConditionModel, DDIMScheduler,
                       DDPMScheduler, AutoencoderKL)

from .mel import Mel

VERSION = "1.2.5"


class AudioDiffusion:

    def __init__(self,
                 model_id: str = "teticio/audio-diffusion-256",
                 sample_rate: int = 22050,
                 n_fft: int = 2048,
                 hop_length: int = 512,
                 top_db: int = 80,
                 cuda: bool = torch.cuda.is_available(),
                 progress_bar: Iterable = tqdm):
        """Class for generating audio using De-noising Diffusion Probabilistic Models.

        Args:
            model_id (String): name of model (local directory or Hugging Face Hub)
            sample_rate (int): sample rate of audio
            n_fft (int): number of Fast Fourier Transforms
            hop_length (int): hop length (a higher number is recommended for lower than 256 y_res)
            top_db (int): loudest in decibels
            cuda (bool): use CUDA?
            progress_bar (iterable): iterable callback for progress updates or None
        """
        self.model_id = model_id
        pipeline = {
            'LatentAudioDiffusionPipeline': LatentAudioDiffusionPipeline,
            'AudioDiffusionPipeline': AudioDiffusionPipeline
        }.get(
            DiffusionPipeline.get_config_dict(self.model_id)['_class_name'],
            AudioDiffusionPipeline)
        self.pipe = pipeline.from_pretrained(self.model_id)
        if cuda:
            self.pipe.to("cuda")
        self.progress_bar = progress_bar or (lambda _: _)

        # For backwards compatibility
        sample_size = (self.pipe.unet.sample_size,
                       self.pipe.unet.sample_size) if type(
                           self.pipe.unet.sample_size
                       ) == int else self.pipe.unet.sample_size
        self.mel = Mel(x_res=sample_size[1],
                       y_res=sample_size[0],
                       sample_rate=sample_rate,
                       n_fft=n_fft,
                       hop_length=hop_length,
                       top_db=top_db)

    def generate_spectrogram_and_audio(
        self,
        steps: int = None,
        generator: torch.Generator = None,
        step_generator: torch.Generator = None,
        eta: float = 0,
        noise: torch.Tensor = None
    ) -> Tuple[Image.Image, Tuple[int, np.ndarray]]:
        """Generate random mel spectrogram and convert to audio.

        Args:
            steps (int): number of de-noising steps (defaults to 50 for DDIM, 1000 for DDPM)
            generator (torch.Generator): random number generator or None
            step_generator (torch.Generator): random number generator used to de-noise or None
            eta (float): parameter between 0 and 1 used with DDIM scheduler
            noise (torch.Tensor): noisy image or None

        Returns:
            PIL Image: mel spectrogram
            (float, np.ndarray): sample rate and raw audio
        """
        images, (sample_rate,
                 audios) = self.pipe(mel=self.mel,
                                     batch_size=1,
                                     steps=steps,
                                     generator=generator,
                                     step_generator=step_generator,
                                     eta=eta,
                                     noise=noise)
        return images[0], (sample_rate, audios[0])

    def generate_spectrogram_and_audio_from_audio(
        self,
        audio_file: str = None,
        raw_audio: np.ndarray = None,
        slice: int = 0,
        start_step: int = 0,
        steps: int = None,
        generator: torch.Generator = None,
        mask_start_secs: float = 0,
        mask_end_secs: float = 0,
        step_generator: torch.Generator = None,
        eta: float = 0,
        noise: torch.Tensor = None
    ) -> Tuple[Image.Image, Tuple[int, np.ndarray]]:
        """Generate random mel spectrogram from audio input and convert to audio.

        Args:
            audio_file (str): must be a file on disk due to Librosa limitation or
            raw_audio (np.ndarray): audio as numpy array
            slice (int): slice number of audio to convert
            start_step (int): step to start from
            steps (int): number of de-noising steps (defaults to 50 for DDIM, 1000 for DDPM)
            generator (torch.Generator): random number generator or None
            mask_start_secs (float): number of seconds of audio to mask (not generate) at start
            mask_end_secs (float): number of seconds of audio to mask (not generate) at end
            step_generator (torch.Generator): random number generator used to de-noise or None
            eta (float): parameter between 0 and 1 used with DDIM scheduler
            noise (torch.Tensor): noisy image or None

        Returns:
            PIL Image: mel spectrogram
            (float, np.ndarray): sample rate and raw audio
        """

        images, (sample_rate,
                 audios) = self.pipe(mel=self.mel,
                                     batch_size=1,
                                     audio_file=audio_file,
                                     raw_audio=raw_audio,
                                     slice=slice,
                                     start_step=start_step,
                                     steps=steps,
                                     generator=generator,
                                     mask_start_secs=mask_start_secs,
                                     mask_end_secs=mask_end_secs,
                                     step_generator=step_generator,
                                     eta=eta,
                                     noise=noise)
        return images[0], (sample_rate, audios[0])

    @staticmethod
    def loop_it(audio: np.ndarray,
                sample_rate: int,
                loops: int = 12) -> np.ndarray:
        """Loop audio

        Args:
            audio (np.ndarray): audio as numpy array
            sample_rate (int): sample rate of audio
            loops (int): number of times to loop

        Returns:
            (float, np.ndarray): sample rate and raw audio or None
        """
        _, beats = beat_track(y=audio, sr=sample_rate, units='samples')
        for beats_in_bar in [16, 12, 8, 4]:
            if len(beats) > beats_in_bar:
                return np.tile(audio[beats[0]:beats[beats_in_bar]], loops)
        return None


class AudioDiffusionPipeline(DiffusionPipeline):

    def __init__(self, unet: UNet2DConditionModel,
                 scheduler: Union[DDIMScheduler, DDPMScheduler]):
        super().__init__()
        self.register_modules(unet=unet, scheduler=scheduler)

    @torch.no_grad()
    def __call__(
        self,
        mel: Mel,
        batch_size: int = 1,
        audio_file: str = None,
        raw_audio: np.ndarray = None,
        slice: int = 0,
        start_step: int = 0,
        steps: int = None,
        generator: torch.Generator = None,
        mask_start_secs: float = 0,
        mask_end_secs: float = 0,
        step_generator: torch.Generator = None,
        eta: float = 0,
        noise: torch.Tensor = None
    ) -> Tuple[List[Image.Image], Tuple[int, List[np.ndarray]]]:
        """Generate random mel spectrogram from audio input and convert to audio.

        Args:
            mel (Mel): instance of Mel class to perform image <-> audio
            batch_size (int): number of samples to generate
            audio_file (str): must be a file on disk due to Librosa limitation or
            raw_audio (np.ndarray): audio as numpy array
            slice (int): slice number of audio to convert
            start_step (int): step to start from
            steps (int): number of de-noising steps (defaults to 50 for DDIM, 1000 for DDPM)
            generator (torch.Generator): random number generator or None
            mask_start_secs (float): number of seconds of audio to mask (not generate) at start
            mask_end_secs (float): number of seconds of audio to mask (not generate) at end
            step_generator (torch.Generator): random number generator used to de-noise or None
            eta (float): parameter between 0 and 1 used with DDIM scheduler
            noise (torch.Tensor): noise tensor of shape (batch_size, 1, height, width) or None

        Returns:
            List[PIL Image]: mel spectrograms
            (float, List[np.ndarray]): sample rate and raw audios
        """

        steps = steps or 50 if isinstance(self.scheduler,
                                          DDIMScheduler) else 1000
        self.scheduler.set_timesteps(steps)
        step_generator = step_generator or generator
        # For backwards compatibility
        if type(self.unet.sample_size) == int:
            self.unet.sample_size = (self.unet.sample_size,
                                     self.unet.sample_size)
        if noise is None:
            noise = torch.randn(
                (batch_size, self.unet.in_channels, self.unet.sample_size[0],
                 self.unet.sample_size[1]),
                generator=generator)
        images = noise
        mask = None

        if audio_file is not None or raw_audio is not None:
            mel.load_audio(audio_file, raw_audio)
            input_image = mel.audio_slice_to_image(slice)
            input_image = np.frombuffer(input_image.tobytes(),
                                        dtype="uint8").reshape(
                                            (input_image.height,
                                             input_image.width))
            input_image = ((input_image / 255) * 2 - 1)
            input_images = np.tile(input_image, (batch_size, 1, 1, 1))

            if hasattr(self, 'vqvae'):
                input_images = self.vqvae.encode(
                    input_images).latent_dist.sample(generator=generator)
                input_images = 0.18215 * input_images

            if start_step > 0:
                images[0, 0] = self.scheduler.add_noise(
                    torch.tensor(input_images[:, np.newaxis, np.newaxis, :]),
                    noise, torch.tensor(steps - start_step))

            pixels_per_second = (self.unet.sample_size[1] *
                                 mel.get_sample_rate() / mel.x_res /
                                 mel.hop_length)
            mask_start = int(mask_start_secs * pixels_per_second)
            mask_end = int(mask_end_secs * pixels_per_second)
            mask = self.scheduler.add_noise(
                torch.tensor(input_images[:, np.newaxis, :]), noise,
                torch.tensor(self.scheduler.timesteps[start_step:]))

        images = images.to(self.device)
        for step, t in enumerate(
                self.progress_bar(self.scheduler.timesteps[start_step:])):
            model_output = self.unet(images, t)['sample']

            if isinstance(self.scheduler, DDIMScheduler):
                images = self.scheduler.step(
                    model_output=model_output,
                    timestep=t,
                    sample=images,
                    eta=eta,
                    generator=step_generator)['prev_sample']
            else:
                images = self.scheduler.step(
                    model_output=model_output,
                    timestep=t,
                    sample=images,
                    generator=step_generator)['prev_sample']

            if mask is not None:
                if mask_start > 0:
                    images[:, :, :, :mask_start] = mask[
                        step, :, :, :, :mask_start]
                if mask_end > 0:
                    images[:, :, :, -mask_end:] = mask[step, :, :, :,
                                                       -mask_end:]

        if hasattr(self, 'vqvae'):
            # 0.18215 was scaling factor used in training to ensure unit variance
            images = 1 / 0.18215 * images
            images = self.vqvae.decode(images)['sample']

        images = (images / 2 + 0.5).clamp(0, 1)
        images = images.cpu().permute(0, 2, 3, 1).numpy()
        images = (images * 255).round().astype("uint8")
        images = list(
            map(lambda _: Image.fromarray(_[:, :, 0]), images) if images.
            shape[3] == 1 else map(
                lambda _: Image.fromarray(_, mode='RGB').convert('L'), images))

        audios = list(map(lambda _: mel.image_to_audio(_), images))
        return images, (mel.get_sample_rate(), audios)

    @torch.no_grad()
    def encode(self, images: List[Image.Image], steps: int = 50) -> np.ndarray:
        """Reverse step process: recover noisy image from generated image.

        Args:
            images (List[PIL Image]): list of images to encode
            steps (int): number of encoding steps to perform (defaults to 50)

        Returns:
            np.ndarray: noise tensor of shape (batch_size, 1, height, width)
        """

        # Only works with DDIM as this method is deterministic
        assert isinstance(self.scheduler, DDIMScheduler)
        self.scheduler.set_timesteps(steps)
        sample = np.array([
            np.frombuffer(image.tobytes(), dtype="uint8").reshape(
                (1, image.height, image.width)) for image in images
        ])
        sample = ((sample / 255) * 2 - 1)
        sample = torch.Tensor(sample).to(self.device)

        for t in self.progress_bar(torch.flip(self.scheduler.timesteps,
                                              (0, ))):
            prev_timestep = (t - self.scheduler.num_train_timesteps //
                             self.scheduler.num_inference_steps)
            alpha_prod_t = self.scheduler.alphas_cumprod[t]
            alpha_prod_t_prev = (self.scheduler.alphas_cumprod[prev_timestep]
                                 if prev_timestep >= 0 else
                                 self.scheduler.final_alpha_cumprod)
            beta_prod_t = 1 - alpha_prod_t
            model_output = self.unet(sample, t)['sample']
            pred_sample_direction = (1 -
                                     alpha_prod_t_prev)**(0.5) * model_output
            sample = (sample -
                      pred_sample_direction) * alpha_prod_t_prev**(-0.5)
            sample = sample * alpha_prod_t**(0.5) + beta_prod_t**(
                0.5) * model_output

        return sample

    @staticmethod
    def slerp(x0: torch.Tensor, x1: torch.Tensor,
              alpha: float) -> torch.Tensor:
        """Spherical Linear intERPolation

        Args:
            x0 (torch.Tensor): first tensor to interpolate between
            x1 (torch.Tensor): seconds tensor to interpolate between
            alpha (float): interpolation between 0 and 1

        Returns:
            torch.Tensor: interpolated tensor
        """

        theta = acos(
            torch.dot(torch.flatten(x0), torch.flatten(x1)) / torch.norm(x0) /
            torch.norm(x1))
        return sin((1 - alpha) * theta) * x0 / sin(theta) + sin(
            alpha * theta) * x1 / sin(theta)


class LatentAudioDiffusionPipeline(AudioDiffusionPipeline):

    def __init__(self, unet: UNet2DConditionModel,
                 scheduler: Union[DDIMScheduler,
                                  DDPMScheduler], vqvae: AutoencoderKL):
        super().__init__(unet=unet, scheduler=scheduler)
        self.register_modules(vqvae=vqvae)

    def __call__(self, *args, **kwargs):
        return super().__call__(*args, **kwargs)