justyoung/DiffSinger
1
1import math
2import random
3from collections import deque
4from functools import partial
5from inspect import isfunction
6from pathlib import Path
7import numpy as np
8import torch
9import torch.nn.functional as F
10from torch import nn
11from tqdm import tqdm
12from einops import rearrange
13
14from modules.fastspeech.fs2 import FastSpeech2
15from modules.diffsinger_midi.fs2 import FastSpeech2MIDI
16from utils.hparams import hparams
17
18
19
20def exists(x):
21 return x is not None
22
23
24def default(val, d):
25 if exists(val):
26 return val
27 return d() if isfunction(d) else d
28
29
30# gaussian diffusion trainer class
31
32def extract(a, t, x_shape):
33 b, *_ = t.shape
34 out = a.gather(-1, t)
35 return out.reshape(b, *((1,) * (len(x_shape) - 1)))
36
37
38def noise_like(shape, device, repeat=False):
39 repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
40 noise = lambda: torch.randn(shape, device=device)
41 return repeat_noise() if repeat else noise()
42
43
44def linear_beta_schedule(timesteps, max_beta=hparams.get('max_beta', 0.01)):
45 """
46 linear schedule
47 """
48 betas = np.linspace(1e-4, max_beta, timesteps)
49 return betas
50
51
52def cosine_beta_schedule(timesteps, s=0.008):
53 """
54 cosine schedule
55 as proposed in https://openreview.net/forum?id=-NEXDKk8gZ
56 """
57 steps = timesteps + 1
58 x = np.linspace(0, steps, steps)
59 alphas_cumprod = np.cos(((x / steps) + s) / (1 + s) * np.pi * 0.5) ** 2
60 alphas_cumprod = alphas_cumprod / alphas_cumprod[0]
61 betas = 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1])
62 return np.clip(betas, a_min=0, a_max=0.999)
63
64
65beta_schedule = {
66 "cosine": cosine_beta_schedule,
67 "linear": linear_beta_schedule,
68}
69
70
71class GaussianDiffusion(nn.Module):
72 def __init__(self, phone_encoder, out_dims, denoise_fn,
73 timesteps=1000, K_step=1000, loss_type=hparams.get('diff_loss_type', 'l1'), betas=None, spec_min=None, spec_max=None):
74 super().__init__()
75 self.denoise_fn = denoise_fn
76 if hparams.get('use_midi') is not None and hparams['use_midi']:
77 self.fs2 = FastSpeech2MIDI(phone_encoder, out_dims)
78 else:
79 self.fs2 = FastSpeech2(phone_encoder, out_dims)
80 self.mel_bins = out_dims
81
82 if exists(betas):
83 betas = betas.detach().cpu().numpy() if isinstance(betas, torch.Tensor) else betas
84 else:
85 if 'schedule_type' in hparams.keys():
86 betas = beta_schedule[hparams['schedule_type']](timesteps)
87 else:
88 betas = cosine_beta_schedule(timesteps)
89
90 alphas = 1. - betas
91 alphas_cumprod = np.cumprod(alphas, axis=0)
92 alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
93
94 timesteps, = betas.shape
95 self.num_timesteps = int(timesteps)
96 self.K_step = K_step
97 self.loss_type = loss_type
98
99 self.noise_list = deque(maxlen=4)
100
101 to_torch = partial(torch.tensor, dtype=torch.float32)
102
103 self.register_buffer('betas', to_torch(betas))
104 self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
105 self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
106
107 # calculations for diffusion q(x_t | x_{t-1}) and others
108 self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
109 self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
110 self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
111 self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
112 self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
113
114 # calculations for posterior q(x_{t-1} | x_t, x_0)
115 posterior_variance = betas * (1. - alphas_cumprod_prev) / (1. - alphas_cumprod)
116 # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
117 self.register_buffer('posterior_variance', to_torch(posterior_variance))
118 # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
119 self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
120 self.register_buffer('posterior_mean_coef1', to_torch(
121 betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
122 self.register_buffer('posterior_mean_coef2', to_torch(
123 (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
124
125 self.register_buffer('spec_min', torch.FloatTensor(spec_min)[None, None, :hparams['keep_bins']])
126 self.register_buffer('spec_max', torch.FloatTensor(spec_max)[None, None, :hparams['keep_bins']])
127
128 def q_mean_variance(self, x_start, t):
129 mean = extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
130 variance = extract(1. - self.alphas_cumprod, t, x_start.shape)
131 log_variance = extract(self.log_one_minus_alphas_cumprod, t, x_start.shape)
132 return mean, variance, log_variance
133
134 def predict_start_from_noise(self, x_t, t, noise):
135 return (
136 extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
137 extract(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
138 )
139
140 def q_posterior(self, x_start, x_t, t):
141 posterior_mean = (
142 extract(self.posterior_mean_coef1, t, x_t.shape) * x_start +
143 extract(self.posterior_mean_coef2, t, x_t.shape) * x_t
144 )
145 posterior_variance = extract(self.posterior_variance, t, x_t.shape)
146 posterior_log_variance_clipped = extract(self.posterior_log_variance_clipped, t, x_t.shape)
147 return posterior_mean, posterior_variance, posterior_log_variance_clipped
148
149 def p_mean_variance(self, x, t, cond, clip_denoised: bool):
150 noise_pred = self.denoise_fn(x, t, cond=cond)
151 x_recon = self.predict_start_from_noise(x, t=t, noise=noise_pred)
152
153 if clip_denoised:
154 x_recon.clamp_(-1., 1.)
155
156 model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
157 return model_mean, posterior_variance, posterior_log_variance
158
159 @torch.no_grad()
160 def p_sample(self, x, t, cond, clip_denoised=True, repeat_noise=False):
161 b, *_, device = *x.shape, x.device
162 model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, cond=cond, clip_denoised=clip_denoised)
163 noise = noise_like(x.shape, device, repeat_noise)
164 # no noise when t == 0
165 nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
166 return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
167
168 @torch.no_grad()
169 def p_sample_plms(self, x, t, interval, cond, clip_denoised=True, repeat_noise=False):
170 """
171 Use the PLMS method from [Pseudo Numerical Methods for Diffusion Models on Manifolds](https://arxiv.org/abs/2202.09778).
172 """
173
174 def get_x_pred(x, noise_t, t):
175 a_t = extract(self.alphas_cumprod, t, x.shape)
176 if t[0] < interval:
177 a_prev = torch.ones_like(a_t)
178 else:
179 a_prev = extract(self.alphas_cumprod, torch.max(t-interval, torch.zeros_like(t)), x.shape)
180 a_t_sq, a_prev_sq = a_t.sqrt(), a_prev.sqrt()
181
182 x_delta = (a_prev - a_t) * ((1 / (a_t_sq * (a_t_sq + a_prev_sq))) * x - 1 / (a_t_sq * (((1 - a_prev) * a_t).sqrt() + ((1 - a_t) * a_prev).sqrt())) * noise_t)
183 x_pred = x + x_delta
184
185 return x_pred
186
187 noise_list = self.noise_list
188 noise_pred = self.denoise_fn(x, t, cond=cond)
189
190 if len(noise_list) == 0:
191 x_pred = get_x_pred(x, noise_pred, t)
192 noise_pred_prev = self.denoise_fn(x_pred, max(t-interval, 0), cond=cond)
193 noise_pred_prime = (noise_pred + noise_pred_prev) / 2
194 elif len(noise_list) == 1:
195 noise_pred_prime = (3 * noise_pred - noise_list[-1]) / 2
196 elif len(noise_list) == 2:
197 noise_pred_prime = (23 * noise_pred - 16 * noise_list[-1] + 5 * noise_list[-2]) / 12
198 elif len(noise_list) >= 3:
199 noise_pred_prime = (55 * noise_pred - 59 * noise_list[-1] + 37 * noise_list[-2] - 9 * noise_list[-3]) / 24
200
201 x_prev = get_x_pred(x, noise_pred_prime, t)
202 noise_list.append(noise_pred)
203
204 return x_prev
205
206 def q_sample(self, x_start, t, noise=None):
207 noise = default(noise, lambda: torch.randn_like(x_start))
208 return (
209 extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
210 extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise
211 )
212
213 def p_losses(self, x_start, t, cond, noise=None, nonpadding=None):
214 noise = default(noise, lambda: torch.randn_like(x_start))
215
216 x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
217 x_recon = self.denoise_fn(x_noisy, t, cond)
218
219 if self.loss_type == 'l1':
220 if nonpadding is not None:
221 loss = ((noise - x_recon).abs() * nonpadding.unsqueeze(1)).mean()
222 else:
223 # print('are you sure w/o nonpadding?')
224 loss = (noise - x_recon).abs().mean()
225
226 elif self.loss_type == 'l2':
227 loss = F.mse_loss(noise, x_recon)
228 else:
229 raise NotImplementedError()
230
231 return loss
232
233 def forward(self, txt_tokens, mel2ph=None, spk_embed=None,
234 ref_mels=None, f0=None, uv=None, energy=None, infer=False, **kwargs):
235 b, *_, device = *txt_tokens.shape, txt_tokens.device
236 ret = self.fs2(txt_tokens, mel2ph, spk_embed, ref_mels, f0, uv, energy,
237 skip_decoder=(not infer), infer=infer, **kwargs)
238 cond = ret['decoder_inp'].transpose(1, 2)
239
240 if not infer:
241 t = torch.randint(0, self.K_step, (b,), device=device).long()
242 x = ref_mels
243 x = self.norm_spec(x)
244 x = x.transpose(1, 2)[:, None, :, :] # [B, 1, M, T]
245 ret['diff_loss'] = self.p_losses(x, t, cond)
246 # nonpadding = (mel2ph != 0).float()
247 # ret['diff_loss'] = self.p_losses(x, t, cond, nonpadding=nonpadding)
248 else:
249 ret['fs2_mel'] = ret['mel_out']
250 fs2_mels = ret['mel_out']
251 t = self.K_step
252 fs2_mels = self.norm_spec(fs2_mels)
253 fs2_mels = fs2_mels.transpose(1, 2)[:, None, :, :]
254
255 x = self.q_sample(x_start=fs2_mels, t=torch.tensor([t - 1], device=device).long())
256 if hparams.get('gaussian_start') is not None and hparams['gaussian_start']:
257 print('===> gaussion start.')
258 shape = (cond.shape[0], 1, self.mel_bins, cond.shape[2])
259 x = torch.randn(shape, device=device)
260
261 if hparams.get('pndm_speedup'):
262 print('===> pndm speed:', hparams['pndm_speedup'])
263 self.noise_list = deque(maxlen=4)
264 iteration_interval = hparams['pndm_speedup']
265 for i in tqdm(reversed(range(0, t, iteration_interval)), desc='sample time step',
266 total=t // iteration_interval):
267 x = self.p_sample_plms(x, torch.full((b,), i, device=device, dtype=torch.long), iteration_interval,
268 cond)
269 else:
270 for i in tqdm(reversed(range(0, t)), desc='sample time step', total=t):
271 x = self.p_sample(x, torch.full((b,), i, device=device, dtype=torch.long), cond)
272 x = x[:, 0].transpose(1, 2)
273 if mel2ph is not None: # for singing
274 ret['mel_out'] = self.denorm_spec(x) * ((mel2ph > 0).float()[:, :, None])
275 else:
276 ret['mel_out'] = self.denorm_spec(x)
277 return ret
278
279 def norm_spec(self, x):
280 return (x - self.spec_min) / (self.spec_max - self.spec_min) * 2 - 1
281
282 def denorm_spec(self, x):
283 return (x + 1) / 2 * (self.spec_max - self.spec_min) + self.spec_min
284
285 def cwt2f0_norm(self, cwt_spec, mean, std, mel2ph):
286 return self.fs2.cwt2f0_norm(cwt_spec, mean, std, mel2ph)
287
288 def out2mel(self, x):
289 return x
290
291
292class OfflineGaussianDiffusion(GaussianDiffusion):
293 def forward(self, txt_tokens, mel2ph=None, spk_embed=None,
294 ref_mels=None, f0=None, uv=None, energy=None, infer=False, **kwargs):
295 b, *_, device = *txt_tokens.shape, txt_tokens.device
296
297 ret = self.fs2(txt_tokens, mel2ph, spk_embed, ref_mels, f0, uv, energy,
298 skip_decoder=True, infer=True, **kwargs)
299 cond = ret['decoder_inp'].transpose(1, 2)
300 fs2_mels = ref_mels[1]
301 ref_mels = ref_mels[0]
302
303 if not infer:
304 t = torch.randint(0, self.K_step, (b,), device=device).long()
305 x = ref_mels
306 x = self.norm_spec(x)
307 x = x.transpose(1, 2)[:, None, :, :] # [B, 1, M, T]
308 ret['diff_loss'] = self.p_losses(x, t, cond)
309 else:
310 t = self.K_step
311 fs2_mels = self.norm_spec(fs2_mels)
312 fs2_mels = fs2_mels.transpose(1, 2)[:, None, :, :]
313
314 x = self.q_sample(x_start=fs2_mels, t=torch.tensor([t - 1], device=device).long())
315
316 if hparams.get('gaussian_start') is not None and hparams['gaussian_start']:
317 print('===> gaussion start.')
318 shape = (cond.shape[0], 1, self.mel_bins, cond.shape[2])
319 x = torch.randn(shape, device=device)
320 for i in tqdm(reversed(range(0, t)), desc='sample time step', total=t):
321 x = self.p_sample(x, torch.full((b,), i, device=device, dtype=torch.long), cond)
322 x = x[:, 0].transpose(1, 2)
323 ret['mel_out'] = self.denorm_spec(x)
324 return ret
325