File size: 15,866 Bytes
64e7f2f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
import os
import numpy as np
import torch
import torch.nn as nn
import copy
from tqdm import tqdm
def flatten(v):
    """
    Flatten a list of lists/tuples
    """

    return [x for y in v for x in y]


def rescale(x):
    """
    Rescale a tensor to 0-1
    """

    return (x - x.min()) / (x.max() - x.min())


def find_max_epoch(path):
    """
    Find maximum epoch/iteration in path, formatted ${n_iter}.pkl
    E.g. 100000.pkl

    Parameters:
    path (str): checkpoint path
    
    Returns:
    maximum iteration, -1 if there is no (valid) checkpoint
    """

    files = os.listdir(path)
    epoch = -1
    for f in files:
        if len(f) <= 4:
            continue
        if f[-4:]  == '.pkl':
            try:
                epoch = max(epoch, int(f[:-4]))
            except:
                continue
    #print(path, epoch, flush=True)
    return epoch


def print_size(net):
    """
    Print the number of parameters of a network
    """

    if net is not None and isinstance(net, torch.nn.Module):
        module_parameters = filter(lambda p: p.requires_grad, net.parameters())
        params = sum([np.prod(p.size()) for p in module_parameters])
        print("{} Parameters: {:.6f}M".format(
            net.__class__.__name__, params / 1e6), flush=True)


# Utilities for diffusion models

def std_normal(size):
    """
    Generate the standard Gaussian variable of a certain size
    """

    return torch.normal(0, 1, size=size)


def calc_noise_scale_embedding(noise_scales, noise_scale_embed_dim_in):
    """
    Embed a noise scale $t$ into a higher dimensional space
    E.g. the embedding vector in the 128-dimensional space is
    [sin(t * 10^(0*4/63)), ... , sin(t * 10^(63*4/63)), cos(t * 10^(0*4/63)), ... , cos(t * 10^(63*4/63))]

    Parameters:
    noise_scales (torch.long tensor, shape=(batchsize, 1)):     
                                noise scales for batch data
    noise_scale_embed_dim_in (int, default=128):  
                                dimensionality of the embedding space for discrete noise scales
    
    Returns:
    the embedding vectors (torch.tensor, shape=(batchsize, noise_scale_embed_dim_in)):
    """

    assert noise_scale_embed_dim_in % 2 == 0

    half_dim = noise_scale_embed_dim_in // 2
    _embed = np.log(10000) / (half_dim - 1)
    _embed = torch.exp(torch.arange(half_dim) * -_embed)
    _embed = noise_scales * _embed
    noise_scale_embed = torch.cat((torch.sin(_embed), 
                                      torch.cos(_embed)), 1)
    
    return noise_scale_embed


def calc_diffusion_hyperparams_given_beta(beta):
    """
    Compute diffusion process hyperparameters

    Parameters:
    beta (tensor):  beta schedule 
    
    Returns:
    a dictionary of diffusion hyperparameters including:
        T (int), beta/alpha/sigma (torch.tensor on cpu, shape=(T, ))
        These cpu tensors are changed to cuda tensors on each individual gpu
    """

    T = len(beta)
    alpha = 1 - beta
    sigma = beta + 0
    for t in range(1, T):
        alpha[t] *= alpha[t-1]  # \alpha^2_t = \prod_{s=1}^t (1-\beta_s)
        sigma[t] *= (1-alpha[t-1]) / (1-alpha[t])  # \sigma^2_t = \beta_t * (1-\alpha_{t-1}) / (1-\alpha_t)
    alpha = torch.sqrt(alpha)
    sigma = torch.sqrt(sigma)
    
    _dh = {}
    _dh["T"], _dh["beta"], _dh["alpha"], _dh["sigma"] = T, beta, alpha, sigma
    diffusion_hyperparams = _dh
    return diffusion_hyperparams


def calc_diffusion_hyperparams(T, beta_0, beta_T, tau, N, beta_N, alpha_N, rho):
    """
    Compute diffusion process hyperparameters

    Parameters:
    T (int):                    number of noise scales
    beta_0 and beta_T (float):  beta schedule start/end value, 
                                where any beta_t in the middle is linearly interpolated
    
    Returns:
    a dictionary of diffusion hyperparameters including:
        T (int), beta/alpha/sigma (torch.tensor on cpu, shape=(T, ))
        These cpu tensors are changed to cuda tensors on each individual gpu
    """

    beta = torch.linspace(beta_0, beta_T, T)
    alpha = 1 - beta
    sigma = beta + 0
    for t in range(1, T):
        alpha[t] *= alpha[t-1]  # \alpha^2_t = \prod_{s=1}^t (1-\beta_s)
        sigma[t] *= (1-alpha[t-1]) / (1-alpha[t])  # \sigma^2_t = \beta_t * (1-\alpha_{t-1}) / (1-\alpha_t)
    alpha = torch.sqrt(alpha)
    sigma = torch.sqrt(sigma)
    
    _dh = {}
    _dh["T"], _dh["beta"], _dh["alpha"], _dh["sigma"] = T, beta, alpha, sigma
    _dh["tau"], _dh["N"], _dh["betaN"], _dh["alphaN"], _dh["rho"] = tau, N, beta_N, alpha_N, rho
    diffusion_hyperparams = _dh
    return diffusion_hyperparams


def sampling_given_noise_schedule(
        net,
        size,
        diffusion_hyperparams,
        inference_noise_schedule,
        condition=None,
        ddim=False,
        return_sequence=False):
    """
    Perform the complete sampling step according to p(x_0|x_T) = \prod_{t=1}^T p_{\theta}(x_{t-1}|x_t)

    Parameters:
    net (torch network):            the wavenet models
    size (tuple):                   size of tensor to be generated,
                                    usually is (number of audios to generate, channels=1, length of audio)
    diffusion_hyperparams (dict):   dictionary of diffusion hyperparameters returned by calc_diffusion_hyperparams
                                    note, the tensors need to be cuda tensors
    condition (torch.tensor):       ground truth mel spectrogram read from disk
                                    None if used for unconditional generation

    Returns:
    the generated audio(s) in torch.tensor, shape=size
    """

    _dh = diffusion_hyperparams
    T, alpha = _dh["T"], _dh["alpha"]
    assert len(alpha) == T
    assert len(size) == 3

    N = len(inference_noise_schedule)
    beta_infer = inference_noise_schedule
    alpha_infer = 1 - beta_infer
    sigma_infer = beta_infer + 0
    for n in range(1, N):
        alpha_infer[n] *= alpha_infer[n - 1]
        sigma_infer[n] *= (1 - alpha_infer[n - 1]) / (1 - alpha_infer[n])
    alpha_infer = torch.sqrt(alpha_infer)
    sigma_infer = torch.sqrt(sigma_infer)

    # Mapping noise scales to time steps
    steps_infer = []
    for n in range(N):
        step = map_noise_scale_to_time_step(alpha_infer[n], alpha)
        if step >= 0:
            steps_infer.append(step)
    steps_infer = torch.FloatTensor(steps_infer)

    # N may change since alpha_infer can be out of the range of alpha
    N = len(steps_infer)

    x = std_normal(size)
    if return_sequence:
        x_ = copy.deepcopy(x)
        xs = [x_]
    with torch.no_grad():
        for n in tqdm(range(N - 1, -1, -1), desc='FastDiff sample time step', total=N):
            diffusion_steps = (steps_infer[n] * torch.ones((size[0], 1)))
            epsilon_theta = net((x, condition, diffusion_steps,))
            if ddim:
                alpha_next = alpha_infer[n] / (1 - beta_infer[n]).sqrt()
                c1 = alpha_next / alpha_infer[n]
                c2 = -(1 - alpha_infer[n] ** 2.).sqrt() * c1
                c3 = (1 - alpha_next ** 2.).sqrt()
                x = c1 * x + c2 * epsilon_theta + c3 * epsilon_theta  # std_normal(size)
            else:
                x -= beta_infer[n] / torch.sqrt(1 - alpha_infer[n] ** 2.) * epsilon_theta
                x /= torch.sqrt(1 - beta_infer[n])
                if n > 0:
                    x = x + sigma_infer[n] * std_normal(size)
            if return_sequence:
                x_ = copy.deepcopy(x)
                xs.append(x_)
    if return_sequence:
        return xs
    return x

def noise_scheduling(net, size, diffusion_hyperparams, condition=None, ddim=False):
    """
    Perform the complete sampling step according to p(x_0|x_T) = \prod_{t=1}^T p_{\theta}(x_{t-1}|x_t)

    Parameters:
    net (torch network):            the wavenet models
    size (tuple):                   size of tensor to be generated,
                                    usually is (number of audios to generate, channels=1, length of audio)
    diffusion_hyperparams (dict):   dictionary of diffusion hyperparameters returned by calc_diffusion_hyperparams
                                    note, the tensors need to be cuda tensors
    condition (torch.tensor):       ground truth mel spectrogram read from disk
                                    None if used for unconditional generation

    Returns:
    noise schedule:                 a list of noise scales in torch.tensor, length <= N
    """

    _dh = diffusion_hyperparams
    N, betaN, alphaN, rho, alpha = _dh["N"], _dh["betaN"], _dh["alphaN"], _dh["rho"], _dh["alpha"]

    print('begin noise scheduling, maximum number of reverse steps = %d' % (N))

    betas = []
    x = std_normal(size)
    with torch.no_grad():
        beta_cur = torch.ones(1, 1, 1).cuda() * betaN
        alpha_cur = torch.ones(1, 1, 1).cuda() * alphaN
        for n in range(N - 1, -1, -1):
            # print(n, beta_cur.squeeze().item(), alpha_cur.squeeze().item())
            step = map_noise_scale_to_time_step(alpha_cur.squeeze().item(), alpha)
            if step >= 0:
                betas.append(beta_cur.squeeze().item())
            diffusion_steps = (step * torch.ones((size[0], 1))).cuda()
            epsilon_theta = net((x, condition, diffusion_steps,))
            if ddim:
                alpha_nxt = alpha_cur / (1 - beta_cur).sqrt()
                c1 = alpha_nxt / alpha_cur
                c2 = -(1 - alpha_cur ** 2.).sqrt() * c1
                c3 = (1 - alpha_nxt ** 2.).sqrt()
                x = c1 * x + c2 * epsilon_theta + c3 * epsilon_theta  # std_normal(size)
            else:
                x -= beta_cur / torch.sqrt(1 - alpha_cur ** 2.) * epsilon_theta
                x /= torch.sqrt(1 - beta_cur)
            alpha_nxt, beta_nxt = alpha_cur, beta_cur
            alpha_cur = alpha_nxt / (1 - beta_nxt).sqrt()
            if alpha_cur > 1:
                break
            beta_cur = net.noise_pred(
                x.squeeze(1), (beta_nxt.view(-1, 1), (1 - alpha_cur ** 2.).view(-1, 1)))
            if beta_cur.squeeze().item() < rho:
                break
    return torch.FloatTensor(betas[::-1]).cuda()


def theta_timestep_loss(net, X, diffusion_hyperparams, reverse=False):
    """
    Compute the training loss for learning theta

    Parameters:
    net (torch network):            the wavenet models
    X (tuple, shape=(2,)):          training data in tuple form (mel_spectrograms, audios)
                                    mel_spectrograms: torch.tensor, shape is batchsize followed by each mel_spectrogram shape
                                    audios: torch.tensor, shape=(batchsize, 1, length of audio)
    diffusion_hyperparams (dict):   dictionary of diffusion hyperparameters returned by calc_diffusion_hyperparams
                                    note, the tensors need to be cuda tensors

    Returns:
    theta loss
    """
    assert type(X) == tuple and len(X) == 2
    loss_fn = nn.MSELoss()

    _dh = diffusion_hyperparams
    T, alpha = _dh["T"], _dh["alpha"]

    mel_spectrogram, audio = X
    B, C, L = audio.shape  # B is batchsize, C=1, L is audio length
    ts = torch.randint(T, size=(B, 1, 1)).cuda()  # randomly sample steps from 1~T
    z = std_normal(audio.shape)
    delta = (1 - alpha[ts] ** 2.).sqrt()
    alpha_cur = alpha[ts]
    noisy_audio = alpha_cur * audio + delta * z  # compute x_t from q(x_t|x_0)
    epsilon_theta = net((noisy_audio, mel_spectrogram, ts.view(B, 1),))

    if reverse:
        x0 = (noisy_audio - delta * epsilon_theta) / alpha_cur
        return loss_fn(epsilon_theta, z), x0

    return loss_fn(epsilon_theta, z)


def phi_loss(net, X, diffusion_hyperparams):
    """
    Compute the training loss for learning phi
    Parameters:
    net (torch network):            the wavenet models
    X (tuple, shape=(2,)):          training data in tuple form (mel_spectrograms, audios)
                                    mel_spectrograms: torch.tensor, shape is batchsize followed by each mel_spectrogram shape
                                    audios: torch.tensor, shape=(batchsize, 1, length of audio)
    diffusion_hyperparams (dict):   dictionary of diffusion hyperparameters returned by calc_diffusion_hyperparams
                                    note, the tensors need to be cuda tensors

    Returns:
    phi loss
    """
    assert type(X) == tuple and len(X) == 2
    _dh = diffusion_hyperparams
    T, alpha, tau = _dh["T"], _dh["alpha"], _dh["tau"]

    mel_spectrogram, audio = X
    B, C, L = audio.shape  # B is batchsize, C=1, L is audio length
    ts = torch.randint(tau, T - tau, size=(B,)).cuda()  # randomly sample steps from 1~T
    alpha_cur = alpha.index_select(0, ts).view(B, 1, 1)
    alpha_nxt = alpha.index_select(0, ts + tau).view(B, 1, 1)
    beta_nxt = 1 - (alpha_nxt / alpha_cur) ** 2.
    delta = (1 - alpha_cur ** 2.).sqrt()
    z = std_normal(audio.shape)
    noisy_audio = alpha_cur * audio + delta * z  # compute x_t from q(x_t|x_0)
    epsilon_theta = net((noisy_audio, mel_spectrogram, ts.view(B, 1),))
    beta_est = net.noise_pred(noisy_audio.squeeze(1), (beta_nxt.view(B, 1), delta.view(B, 1) ** 2.))
    phi_loss = 1 / (2. * (delta ** 2. - beta_est)) * (
            delta * z - beta_est / delta * epsilon_theta) ** 2.
    phi_loss += torch.log(1e-8 + delta ** 2. / (beta_est + 1e-8)) / 4.
    phi_loss = (torch.mean(phi_loss, -1, keepdim=True) + beta_est / delta ** 2 / 2.).mean()

    return phi_loss


def compute_hyperparams_given_schedule(beta):
    """
    Compute diffusion process hyperparameters

    Parameters:
    beta (tensor):  beta schedule

    Returns:
    a dictionary of diffusion hyperparameters including:
        T (int), beta/alpha/sigma (torch.tensor on cpu, shape=(T, ))
        These cpu tensors are changed to cuda tensors on each individual gpu
    """

    T = len(beta)
    alpha = 1 - beta
    sigma = beta + 0
    for t in range(1, T):
        alpha[t] *= alpha[t - 1]  # \alpha^2_t = \prod_{s=1}^t (1-\beta_s)
        sigma[t] *= (1 - alpha[t - 1]) / (1 - alpha[t])  # \sigma^2_t = \beta_t * (1-\alpha_{t-1}) / (1-\alpha_t)
    alpha = torch.sqrt(alpha)
    sigma = torch.sqrt(sigma)

    _dh = {}
    _dh["T"], _dh["beta"], _dh["alpha"], _dh["sigma"] = T, beta, alpha, sigma
    diffusion_hyperparams = _dh
    return diffusion_hyperparams



def map_noise_scale_to_time_step(alpha_infer, alpha):
    if alpha_infer < alpha[-1]:
        return len(alpha) - 1
    if alpha_infer > alpha[0]:
        return 0
    for t in range(len(alpha) - 1):
        if alpha[t+1] <= alpha_infer <= alpha[t]:
             step_diff = alpha[t] - alpha_infer
             step_diff /= alpha[t] - alpha[t+1]
             return t + step_diff.item()
    return -1


def calc_diffusion_step_embedding(diffusion_steps, diffusion_step_embed_dim_in):
    """
    Embed a diffusion step $t$ into a higher dimensional space
    E.g. the embedding vector in the 128-dimensional space is
    [sin(t * 10^(0*4/63)), ... , sin(t * 10^(63*4/63)), cos(t * 10^(0*4/63)), ... , cos(t * 10^(63*4/63))]

    Parameters:
    diffusion_steps (torch.long tensor, shape=(batchsize, 1)):
                                diffusion steps for batch data
    diffusion_step_embed_dim_in (int, default=128):
                                dimensionality of the embedding space for discrete diffusion steps

    Returns:
    the embedding vectors (torch.tensor, shape=(batchsize, diffusion_step_embed_dim_in)):
    """

    assert diffusion_step_embed_dim_in % 2 == 0

    half_dim = diffusion_step_embed_dim_in // 2
    _embed = np.log(10000) / (half_dim - 1)
    _embed = torch.exp(torch.arange(half_dim) * -_embed)
    _embed = diffusion_steps * _embed
    diffusion_step_embed = torch.cat((torch.sin(_embed),
                                      torch.cos(_embed)), 1)

    return diffusion_step_embed