File size: 37,780 Bytes
fb0011a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 | # models/reinforce_aligner.py
# Core encoder + phoneme-wise (α=2) Reinforce-Aligner with Gaussian upsampling
# Modified to compute rewards from GT durations instead of mel spectrogram loss.
from typing import Callable, Dict, Optional, Tuple
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
def rms_norm_fn(x: torch.Tensor, eps: float = 1e-5) -> torch.Tensor:
return x * torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + eps)
class AdaRMSNorm(nn.Module):
"""
AdaRMSNorm: (1 + gamma) * rms_norm(x) + beta.
Gamma i Beta generowane dynamicznie z wektora stylu.
"""
def __init__(self, style_dim: int, channels: int, eps: float = 1e-5) -> None:
super().__init__()
self.fc = nn.Linear(style_dim, channels * 2)
self.eps = eps
# Inicjalizacja "identity" - gamma i beta bliskie 0 na start
with torch.no_grad():
self.fc.weight.zero_()
self.fc.bias.zero_()
def forward(self, x: torch.Tensor, style: torch.Tensor) -> torch.Tensor:
# x: [B, T, C], style: [B, S]
gb = self.fc(style) # [B, 2C]
gamma, beta = gb.chunk(2, dim=-1)
# Rozszerzenie wymiarów do [B, 1, C] dla broadcastingu po czasie
gamma = gamma.unsqueeze(1).to(dtype=x.dtype)
beta = beta.unsqueeze(1).to(dtype=x.dtype)
x_norm = x * torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + self.eps)
return (1.0 + gamma) * x_norm + beta
class SwiGLUFFN(nn.Module):
def __init__(self, dim: int, expansion: float = 2.0, dropout: float = 0.0) -> None:
super().__init__()
hidden = int(dim * expansion)
self.fc = nn.Linear(dim, hidden * 2)
self.proj = nn.Linear(hidden, dim)
self.dropout = nn.Dropout(dropout)
def forward(self, x: torch.Tensor) -> torch.Tensor:
a, b = self.fc(x).chunk(2, dim=-1)
x = F.silu(b) * a
x = self.proj(self.dropout(x))
return x
# -------------------- RoPE & Attention --------------------
class RotaryEmbedding(nn.Module):
def __init__(self, dim: int, base: float = 10000.0) -> None:
super().__init__()
if dim % 2 != 0:
raise ValueError("RoPE head_dim must be even.")
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
self.register_buffer("inv_freq", inv_freq, persistent=False)
def get_cos_sin(
self, seq_len: int, device: torch.device, dtype: torch.dtype
) -> Tuple[torch.Tensor, torch.Tensor]:
t = torch.arange(seq_len, device=device, dtype=dtype)
freqs = torch.outer(t, self.inv_freq.to(dtype))
emb = torch.cat([freqs, freqs], dim=-1)
cos = emb.cos().unsqueeze(0).unsqueeze(0)
sin = emb.sin().unsqueeze(0).unsqueeze(0)
return cos, sin
def apply_rope(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor:
x_even = x[..., ::2]
x_odd = x[..., 1::2]
x_rot = torch.stack([-x_odd, x_even], dim=-1).flatten(-2)
return (x * cos) + (x_rot * sin)
class MultiheadSelfAttentionRoPE(nn.Module):
def __init__(
self,
embed_dim: int,
num_heads: int,
dropout: float = 0.0,
use_rope: bool = True,
rope_base: float = 10000.0,
) -> None:
super().__init__()
self.embed_dim = embed_dim
self.num_heads = num_heads
self.head_dim = embed_dim // num_heads
self.scale = 1.0 / math.sqrt(self.head_dim)
self.qkv = nn.Linear(embed_dim, embed_dim * 3)
self.out_proj = nn.Linear(embed_dim, embed_dim)
self.attn_dropout = nn.Dropout(dropout)
self.use_rope = use_rope
self.rope = RotaryEmbedding(self.head_dim, base=rope_base) if use_rope else None
def forward(
self, x: torch.Tensor, key_padding_mask: Optional[torch.Tensor] = None
) -> torch.Tensor:
B, T, C = x.shape
qkv = self.qkv(x)
q, k, v = qkv.chunk(3, dim=-1)
# [B, T, 3C] -> [B, H, T, D]
q = q.view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
k = k.view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
v = v.view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
if self.use_rope:
cos, sin = self.rope.get_cos_sin(T, x.device, x.dtype)
q = apply_rope(q, cos, sin)
k = apply_rope(k, cos, sin)
attn_scores = torch.matmul(q, k.transpose(-2, -1)) * self.scale
if key_padding_mask is not None:
# Mask format: [B, T] -> [B, 1, 1, T] for broadcasting
mask = key_padding_mask.unsqueeze(1).unsqueeze(2)
attn_scores = attn_scores.masked_fill(mask, torch.finfo(attn_scores.dtype).min)
attn_weights = F.softmax(attn_scores, dim=-1)
attn_weights = self.attn_dropout(attn_weights)
out = torch.matmul(attn_weights, v)
out = out.transpose(1, 2).contiguous().view(B, T, C)
return self.out_proj(out)
# -------------------- Reasoning Components --------------------
class ReasoningBlock(nn.Module):
def __init__(
self,
dim: int,
num_heads: int,
expansion: float,
style_dim: int,
attn_dropout: float = 0.0,
resid_dropout: float = 0.0,
rms_eps: float = 1e-5,
use_rope: bool = True,
rope_base: float = 10000.0,
) -> None:
super().__init__()
self.attn = MultiheadSelfAttentionRoPE(
embed_dim=dim,
num_heads=num_heads,
dropout=attn_dropout,
use_rope=use_rope,
rope_base=rope_base,
)
self.ff = SwiGLUFFN(dim=dim, expansion=expansion, dropout=resid_dropout)
self.resid_dropout = nn.Dropout(resid_dropout)
# Stylizacja poprzez AdaRMSNorm
self.ada1 = AdaRMSNorm(style_dim=style_dim, channels=dim, eps=rms_eps)
self.ada2 = AdaRMSNorm(style_dim=style_dim, channels=dim, eps=rms_eps)
def forward(
self,
hidden_states: torch.Tensor,
style: torch.Tensor,
key_padding_mask: Optional[torch.Tensor],
) -> torch.Tensor:
# 1. Attention Block z AdaNorm
attn_out = self.attn(hidden_states, key_padding_mask=key_padding_mask)
hidden_states = self.ada1(
hidden_states + self.resid_dropout(attn_out), style
)
# 2. FFN Block z AdaNorm
mlp_out = self.ff(hidden_states)
hidden_states = self.ada2(
hidden_states + self.resid_dropout(mlp_out), style
)
return hidden_states
class ReasoningModule(nn.Module):
def __init__(self, layers: nn.ModuleList) -> None:
super().__init__()
self.layers = layers
def forward(
self,
hidden_states: torch.Tensor,
input_injection: torch.Tensor,
style: torch.Tensor,
key_padding_mask: Optional[torch.Tensor],
) -> torch.Tensor:
# Standardowa architektura rekurencyjna często dodaje wejście w każdym kroku (Input Injection)
# zapobiega to "zapominaniu" promptu/kontekstu.
hidden_states = hidden_states + input_injection
for layer in self.layers:
hidden_states = layer(
hidden_states=hidden_states,
style=style,
key_padding_mask=key_padding_mask,
)
return hidden_states
def sinusoidal_positional_encoding(
seq_len: int, dim: int, device: torch.device, dtype: torch.dtype
) -> torch.Tensor:
pos = torch.arange(seq_len, device=device, dtype=dtype).unsqueeze(1)
i = torch.arange(dim, device=device, dtype=dtype).unsqueeze(0)
rates = torch.pow(10000.0, -(2 * torch.div(i, 2, rounding_mode="floor")) / dim)
angles = pos * rates
pe = torch.zeros((seq_len, dim), device=device, dtype=dtype)
pe[:, 0::2] = torch.sin(angles[:, 0::2])
pe[:, 1::2] = torch.cos(angles[:, 1::2])
return pe.unsqueeze(0)
# -------------------- FULL HRM ENCODER --------------------
class HRMDurationEncoder(nn.Module):
"""
Pełna implementacja HRM (Hierarchical Reasoning Model).
Zmiany względem wersji uproszczonej:
1. Usunięto 'torch.no_grad()' w pętli -> Pełne Backpropagation Through Time (BPTT).
2. Dodano Step Embeddings dla poziomów H i L -> Model wie, w której iteracji myślenia jest.
3. Zachowano architekturę System 1 / System 2 (Hierarchical Recurrence).
"""
def __init__(
self,
sty_dim: int,
d_model: int,
nlayers: int,
max_dur: int,
dropout: float = 0.1,
num_heads: int = 8,
expansion: float = 2.0,
H_cycles: int = 1, # Liczba cykli "wolnych" (High-level)
L_cycles: int = 2, # Liczba cykli "szybkich" (Low-level) wewnątrz jednego H
rms_norm_eps: float = 1e-5,
use_posenc: bool = False,
use_rope: bool = True,
rope_base: float = 10000.0,
vocab_size: Optional[int] = None,
pad_token_id: Optional[int] = None,
token_emb_dim: Optional[int] = None,
token_emb_dropout: float = 0.0,
token_embedding: Optional[nn.Embedding] = None,
token_proj: Optional[nn.Module] = None,
) -> None:
super().__init__()
self.d_model = d_model
self.sty_dim = sty_dim
self.H_cycles = H_cycles
self.L_cycles = L_cycles
self.use_posenc = use_posenc
# --- Embeddingi ---
self.pad_token_id = pad_token_id
if token_embedding is not None:
self.token_embedding = token_embedding
emb_dim = token_embedding.embedding_dim
elif vocab_size is not None:
emb_dim = token_emb_dim or d_model
self.token_embedding = nn.Embedding(vocab_size, emb_dim, padding_idx=pad_token_id)
else:
self.token_embedding = None
emb_dim = d_model # Zakładamy wejście continuous
self.token_emb_dropout = nn.Dropout(token_emb_dropout)
# Projekcja wejścia (jeśli dim się nie zgadza)
if emb_dim != d_model:
self.token_proj = nn.Linear(emb_dim, d_model)
else:
self.token_proj = nn.Identity()
# Input Projection: łączy embedding treści z embeddingiem stylu
self.in_proj = nn.Linear(d_model + sty_dim, d_model)
self.dropout = nn.Dropout(dropout)
# --- HRM Core ---
# Inicjalizacja stanów ukrytych (parametry uczące się)
self.H_init = nn.Parameter(torch.zeros(1, 1, d_model))
self.L_init = nn.Parameter(torch.zeros(1, 1, d_model))
nn.init.trunc_normal_(self.H_init, std=0.02)
nn.init.trunc_normal_(self.L_init, std=0.02)
# Step Embeddings - KLUCZOWE dla pełnego HRM
# Pozwalają modelowi odróżnić pierwszy krok rozumowania od ostatniego
self.h_step_emb = nn.Embedding(H_cycles, d_model)
self.l_step_emb = nn.Embedding(L_cycles, d_model)
# Moduły rezonowania (Recurrent Bodies)
# Zauważ: ReasoningModule zawiera listę ReasoningBlocks (Stack Transformerów)
self.H_level = ReasoningModule(nn.ModuleList([
ReasoningBlock(
dim=d_model, num_heads=num_heads, expansion=expansion,
style_dim=sty_dim, attn_dropout=dropout, resid_dropout=dropout,
rms_eps=rms_norm_eps, use_rope=use_rope, rope_base=rope_base
) for _ in range(nlayers)
]))
self.L_level = ReasoningModule(nn.ModuleList([
ReasoningBlock(
dim=d_model, num_heads=num_heads, expansion=expansion,
style_dim=sty_dim, attn_dropout=dropout, resid_dropout=dropout,
rms_eps=rms_norm_eps, use_rope=use_rope, rope_base=rope_base
) for _ in range(nlayers)
]))
# --- Output ---
self.dur_head = nn.Linear(d_model, max_dur)
def _prepare_input(self, x, m, device):
"""Obsługa konwersji ID -> Embedding i maskowania."""
# Sprawdzenie czy x to [B, T] (tokeny) czy [B, T, C] (embeddingi)
if x.dtype in [torch.long, torch.int] and x.dim() == 2:
if self.token_embedding is None:
raise ValueError("Otrzymano token IDs, ale brak token_embedding.")
# Automatyczna detekcja maski jeśli nie podana
if m is None and self.pad_token_id is not None:
m = (x == self.pad_token_id)
x = self.token_embedding(x)
x = self.token_emb_dropout(x)
x = self.token_proj(x)
else:
# Zakładamy continuous [B, T, C]
if x.dim() == 3 and x.size(1) == self.d_model: # [B, C, T] -> [B, T, C]
x = x.transpose(1, 2)
return x, m
def forward(
self,
x: torch.Tensor,
style: torch.Tensor,
text_lengths: Optional[torch.Tensor] = None, # Opcjonalne dla kompatybilności
m: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""
x: Token IDs [B, T] lub Embeddings [B, T, C]
style: Style vector [B, S]
m: Padding mask [B, T] (True = padding)
"""
device = x.device
B = x.shape[0]
# 1. Przygotowanie wejścia
x, m = self._prepare_input(x, m, device)
T = x.shape[1]
# 2. Kondycjonowanie wejścia stylem (Time-Distributed)
s_expanded = style.unsqueeze(1).expand(B, T, -1)
x_in = torch.cat([x, s_expanded], dim=-1)
x_in = self.in_proj(x_in)
x_in = self.dropout(x_in)
# Opcjonalne klasyczne kodowanie pozycyjne (zwykle zbędne przy RoPE)
if self.use_posenc:
pe = sinusoidal_positional_encoding(T, self.d_model, device, x.dtype)
x_in = x_in + pe
# 3. Inicjalizacja stanów HRM
# [1, 1, C] -> [B, T, C]
z_H = self.H_init.to(dtype=x.dtype).expand(B, T, -1)
z_L = self.L_init.to(dtype=x.dtype).expand(B, T, -1)
# Maska dla attention
key_padding_mask = m.to(device) if m is not None else None
# 4. Główna pętla Hierarchical Reasoning (Full BPTT)
# H (Slow) - zewnętrzna pętla
for h in range(self.H_cycles):
# Dodanie embeddingu kroku dla H
h_step = torch.tensor(h, device=device)
h_emb = self.h_step_emb(h_step).view(1, 1, -1)
# Wstrzyknięcie informacji o kroku do stanu H
z_H_curr = z_H + h_emb
# L (Fast) - wewnętrzna pętla
for l in range(self.L_cycles):
l_step = torch.tensor(l, device=device)
l_emb = self.l_step_emb(l_step).view(1, 1, -1)
# Wejście do L to: aktualny stan L + kontekst z H + oryginalny input
# To pozwala L "obrabiać" dane w kontekście wyższego poziomu
current_input_L = z_H_curr + x_in + l_emb
z_L = self.L_level(
hidden_states=z_L,
input_injection=current_input_L,
style=style,
key_padding_mask=key_padding_mask
)
# Opcjonalne czyszczenie paddingu po każdym kroku (dla stabilności)
if key_padding_mask is not None:
z_L = z_L.masked_fill(key_padding_mask.unsqueeze(-1), 0.0)
# Po zakończeniu cykli L, aktualizujemy poziom H
# Wejście do H to: aktualny stan H + wynik przemyśleń L
current_input_H = z_L # + x_in (opcjonalnie, ale L już to widziało)
# Zauważ: H_level w pętli zewnętrznej widzi skumulowany efekt L
if h < self.H_cycles: # Zawsze wykonujemy, warunek logiczny dla jasności
z_H = self.H_level(
hidden_states=z_H,
input_injection=current_input_H,
style=style,
key_padding_mask=key_padding_mask
)
if key_padding_mask is not None:
z_H = z_H.masked_fill(key_padding_mask.unsqueeze(-1), 0.0)
# 5. Zwracamy stan H (high-level representation) lub predykcję trwania
# Oryginalny kod zwracał z_H, więc zachowujemy to (głowicę można użyć na zewnątrz)
return z_H
def length_to_mask_second(lengths: torch.Tensor, max_len: Optional[int] = None) -> torch.Tensor:
# lengths: [B] (int)
max_len = int(max_len or lengths.max().item())
rng = torch.arange(max_len, device=lengths.device).unsqueeze(0) # [1, T]
mask = rng < lengths.unsqueeze(1) # [B, T]
return mask
class ResBlock1D(nn.Module):
def __init__(self, channels: int, kernel_size: int, dilations=(1, 3, 5), dropout=0.0):
super().__init__()
layers = []
for d in dilations:
pad = (kernel_size - 1) // 2 * d
layers += [
nn.LeakyReLU(0.2),
nn.Conv1d(channels, channels, kernel_size, padding=pad, dilation=d),
nn.Dropout(dropout),
]
self.body = nn.Sequential(*layers)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return x + self.body(x)
class MRFEncoder1D(nn.Module):
"""
Multi-Receptive Field stack (inspired by HiFi-GAN MRF).
Input: [B, D_in, N_tokens]
Output: [B, D, N_tokens] with D=out_channels
"""
def __init__(
self,
in_channels: int,
out_channels: int = 512,
num_blocks: int = 4,
kernels=(3, 5, 7),
dilations=(1, 3, 5),
dropout: float = 0.0,
):
super().__init__()
self.in_proj = nn.Conv1d(in_channels, out_channels, kernel_size=1)
blocks = []
for _ in range(num_blocks):
# parallel branches then sum
branches = nn.ModuleList([ResBlock1D(out_channels, k, dilations, dropout) for k in kernels])
blocks.append(branches)
self.blocks = nn.ModuleList(blocks)
def forward(self, x: torch.Tensor, token_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
# x: [B, Din, N]
x = self.in_proj(x) # [B, D, N]
for branches in self.blocks:
y = 0
for b in branches:
y = y + b(x)
x = y / len(branches)
if token_mask is not None:
x = x * token_mask.unsqueeze(1).float()
return x
class DurationPredictor(nn.Module):
"""
1D conv duration predictor over token axis.
Input: [B, D, N]
Output: durations_int [B, N] (positive ints), durations_raw [B, N] (float pre-round)
"""
def __init__(self, channels: int = 512, dropout: float = 0.1):
super().__init__()
self.conv1 = nn.Conv1d(channels, channels, kernel_size=3, padding=1)
self.ln1 = nn.LayerNorm(channels)
self.conv2 = nn.Conv1d(channels, channels, kernel_size=3, padding=1)
self.ln2 = nn.LayerNorm(channels)
self.proj = nn.Conv1d(channels, 1, kernel_size=1)
self.dropout = nn.Dropout(dropout)
def forward(self, x: torch.Tensor, token_mask: torch.Tensor, min_dur: int = 1) -> Tuple[torch.Tensor, torch.Tensor]:
# x: [B, D, N], token_mask: [B, N] (bool)
B, D, N = x.shape
h = self.conv1(x) # [B, D, N]
h = F.relu(self.ln1(h.transpose(1, 2)).transpose(1, 2))
h = self.dropout(h)
h = self.conv2(h)
h = F.relu(self.ln2(h.transpose(1, 2)).transpose(1, 2))
h = self.dropout(h)
raw = self.proj(h).squeeze(1) # [B, N]
log_dur = raw * token_mask.float() #
# positive duration: softplus + min_dur
#dur_float = F.softplus(raw) + (min_dur + 1e-8)
# zero-out padding tokens
dur_float = raw * token_mask.float()
# discrete durations (ints) for upsampling and shift
dur_int = torch.clamp(torch.round(dur_float), min=min_dur).long()
dur_int = dur_int * token_mask.long()
return dur_int, dur_float
def gaussian_upsample(
token_feats: torch.Tensor, # [B, D, N]
durations: torch.Tensor, # [B, N] (int or float) - can be fractional
T: torch.Tensor, # [B] frames per sample
sigma_sq: float = 10.0,
token_mask: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Numerically-stable gaussian upsampling.
Uses a softmax over negative squared distances (log-sum-exp stability).
"""
B, D, N = token_feats.shape
device = token_feats.device
# --- MODIFIED: Handle None token_mask by calling length_to_mask_second ---
if token_mask is None:
# Assume full length N is valid for all samples if mask is missing
full_lengths = torch.full((B,), N, device=device, dtype=torch.long)
token_mask = length_to_mask_second(full_lengths, max_len=N)
# -------------------------------------------------------------------------
T_list = [int(T[b].item()) for b in range(B)]
max_T = max(T_list) if len(T_list) > 0 else 0
out_list = []
weights_list = []
for b in range(B):
Nb = int(token_mask[b].sum().item())
Tb = T_list[b]
if Nb <= 0 or Tb <= 0:
out_list.append(torch.zeros(D, max_T, device=device, dtype=token_feats.dtype))
weights_list.append(torch.zeros(N, max_T, device=device, dtype=token_feats.dtype))
continue
E = token_feats[b, :, :Nb] # [D, Nb]
d = durations[b, :Nb].float() # [Nb]
# avoid exact zeros in denom but do it smoothly with eps
d_sum = d.sum().clamp(min=1e-8)
scale = Tb / d_sum
l = d * scale # [Nb] (fractional frames)
# centers
c = torch.cumsum(l, dim=0) - 0.5 * l # [Nb]
t = torch.arange(Tb, device=device).float().unsqueeze(0) # [1, Tb]
# squared distances
dist2 = (t - c.unsqueeze(1)) ** 2 # [Nb, Tb]
# stable softmax along token axis (dim=0)
logits = (-dist2 / float(sigma_sq)).to(torch.float32) # compute in fp32
w = F.softmax(logits, dim=0).to(dist2.dtype) # [Nb, Tb]
# upsample
F_bt = E @ w # [D, Tb]
# pad to max_T
if Tb < max_T:
F_pad = torch.zeros(D, max_T, device=device, dtype=F_bt.dtype)
F_pad[:, :Tb] = F_bt
F_bt = F_pad
out_list.append(F_bt)
# pack back to [N, max_T] with padding zeros
w_full = torch.zeros(N, max_T, device=device, dtype=w.dtype)
w_full[:Nb, :Tb] = w
weights_list.append(w_full)
frame_feats = torch.stack(out_list, dim=0) # [B, D, max_T]
weights = torch.stack(weights_list, dim=0) # [B, N, max_T]
return frame_feats, weights
def alternating_shift(dur_int: torch.Tensor, alpha: int = 2, token_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
"""
Apply alternating +/- alpha shift to durations while trying to keep sum constant.
dur_int: [B, N] ints
"""
B, N = dur_int.shape
device = dur_int.device
idx = torch.arange(N, device=device)
signs = torch.where((idx % 2) == 0, 1, -1).view(1, N).expand(B, -1) # [B, N]
if token_mask is not None:
signs = signs * token_mask.long()
ds = dur_int.to(torch.long) + alpha * signs
# clip to >=1 on valid tokens
if token_mask is not None:
ds = torch.where(token_mask, torch.clamp(ds, min=1), torch.zeros_like(ds))
else:
ds = torch.clamp(ds, min=1)
# adjust last valid token to keep sum equal (compensate any clipping drift)
if token_mask is not None:
valid_counts = token_mask.sum(dim=1) # [B]
else:
valid_counts = torch.full((B,), N, device=device, dtype=torch.long)
for b in range(B):
Nb = int(valid_counts[b].item())
if Nb <= 0:
continue
diff = int(ds[b, :Nb].sum().item() - dur_int[b, :Nb].sum().item())
if diff != 0:
# subtract diff from the last token (keep >=1)
j = Nb - 1
ds[b, j] = max(int(ds[b, j].item() - diff), 1)
return ds
class MelPerFrame:
"""
Simple per-frame mel extractor returning log-mel for a mono waveform in [-1,1].
Use the same parameters you used to build ground-truth mels if you can.
"""
def __init__(self, sr: int = 44100, n_fft: int = 4096, hop: int = 600, win: int = 2400, n_mels: int = 128):
import torchaudio
self.mel = torchaudio.transforms.MelSpectrogram(
sample_rate=sr, n_fft=n_fft, hop_length=hop, win_length=win, n_mels=n_mels
).to('cuda')
self.eps = 1e-5
@torch.no_grad()
def __call__(self, wav: torch.Tensor) -> torch.Tensor:
# wav: [B, T] float
m = self.mel(wav.float()) # [B, n_mels, T_m]
m = torch.log(m + self.eps)
return m
def per_phoneme_losses_from_frame_losses(
frame_loss: torch.Tensor, # [B, T]
weights: torch.Tensor, # [B, N, T] (may be padded to max_T)
token_mask: torch.Tensor, # [B, N] (bool)
) -> torch.Tensor:
"""
Average frame losses over frames weighted by token responsibilities.
Handles potential mismatch between weights' time dim and frame_loss time dim
by using only the overlapping prefix (min time length).
Returns [B, N], zeros where token has no coverage or token is padding.
"""
B, N, Tw = weights.shape
if frame_loss.dim() != 2 or frame_loss.shape[0] != B:
raise ValueError("frame_loss must be [B, T] with same B as weights")
Tf = frame_loss.shape[1]
T_use = min(Tw, Tf)
if T_use == 0:
return torch.zeros(B, N, device=weights.device, dtype=weights.dtype)
# Trim to overlapping frames
w = weights[..., :T_use] # [B, N, T_use]
fl = frame_loss[..., :T_use] # [B, T_use]
# weighted average per token
num = (w * fl.unsqueeze(1)).sum(dim=2) # [B, N]
den_raw = w.sum(dim=2) # [B, N]
# safe division: only divide where den_raw > 0
den = den_raw + 1e-8
L = num / den
# zero tokens that had no coverage in the overlapped window
L = torch.where(den_raw > 1e-8, L, torch.zeros_like(L))
# zero out padding tokens
L = L * token_mask.float()
return L
class ReinforceAligner(nn.Module):
"""
End-to-end encoder + RL duration aligner with Gaussian upsampling.
Now: uses provided ground-truth durations to compute per-phoneme rewards.
"""
def __init__(
self,
phoneme_dim: int, # input embedding dim if you pass x_emb; or vocab size if you pass token ids
out_channels: int = 512, # encoder output dim (must match your decoder's ASR_down input)
sigma_sq: float = 10.0,
min_dur: int = 1,
num_mrf_blocks: int = 4,
use_input_is_embeds: bool = True, # True if you pass embeddings [B, N, D]; False if you pass token ids
vocab_size: int = 178, # used only if use_input_is_embeds=False (i.e., you pass ids)
dropout: float = 0.1,
):
super().__init__()
self.use_input_is_embeds = use_input_is_embeds
if not use_input_is_embeds:
self.embed = nn.Embedding(vocab_size, phoneme_dim)
self.encoder = MRFEncoder1D(in_channels=phoneme_dim, out_channels=out_channels,
num_blocks=num_mrf_blocks, dropout=dropout)
self.duration_predictor = DurationPredictor(channels=out_channels, dropout=dropout)
self.sigma_sq = sigma_sq
self.min_dur = min_dur
def inference(
self,
x: torch.Tensor,
token_lengths: torch.Tensor,
):
"""
Inference-time upsampling.
Args:
x: [B, D, N] if using embeddings; otherwise token ids (handled by self.use_input_is_embeds).
token_lengths: [B] token counts.
frame_lengths: optional [B] desired output frames length (e.g., mel length). If None,
we use rounded sum of predicted durations.
use_d_gt: optional boolean mask [B] to force using gt durations if you pass them (not used here).
with_weights: return weights matrix if True.
Returns:
dict with keys:
- "enc": encoder features [B, D, N]
- "dur_int": integer durations from predictor [B, N]
- "dur_log": raw predictor output (log space) [B, N]
- "dur": unlogged durations (frames) [B, N] = exp(dur_log)-1
- "frame_feats": upsampled frame features [B, D, T_max]
- "weights": responsibilities [B, N, T_max] (only if with_weights True)
- "token_mask": [B, N] bool
"""
device = x.device
B = x.size(0)
# embeddings if needed
if self.use_input_is_embeds:
x_emb = x # [B, D, N]
else:
x_emb = self.embed(x).transpose(1, 2) # [B, D, N]
token_mask = length_to_mask_second(token_lengths, max_len=x_emb.size(2)) # [B, N]
maskf = token_mask.float()
# encoder
enc = self.encoder(x_emb, token_mask=token_mask) # [B, D, N]
# durations from predictor (raw/log and integer)
dur_int, dur_log = self.duration_predictor(enc, token_mask, min_dur=self.min_dur) # [B, N], [B, N]
dur_log = dur_log * maskf
# convert to linear durations used in forward()
dur_unlogged = (torch.exp(dur_log) - 1.0) * maskf # [B, N]
T_pred = dur_unlogged.sum(dim=1).clamp(min=1.0).round().long().to(device) # [B]
# optionally, if you want to allow forcing GT durations at inference, you can pass them
# via use_d_gt and gt_durations; this example doesn't handle gt_durations here.
frame_feats, weights = gaussian_upsample(
enc,
dur_unlogged,
T_pred,
sigma_sq=self.sigma_sq,
token_mask=token_mask,
)
return frame_feats ,dur_unlogged , T_pred
def forward(
self,
x: torch.Tensor, # [B, D, N] if use_input_is_embeds else [B, N] token ids
token_lengths: torch.Tensor, # [B] (number of tokens per sample)
frame_lengths: torch.Tensor, # [B] (number of frames per sample; e.g., mel length)
gt_durations: Optional[torch.Tensor] = None, # [B, N] ground-truth durations per token (frames)
use_d_gt: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor, Dict]:
'''
if gt_durations is None:
raise ValueError("gt_durations must be provided when using GT-duration rewards.")
device = x.device
B = x.size(0)
# Prepare embeddings & token mask (compute token_mask after embedding so shapes are stable)
if self.use_input_is_embeds:
x_emb = x # [B, D, N]
else:
x_emb = self.embed(x).transpose(1, 2) # [B, D, N]
token_mask = length_to_mask_second(token_lengths, max_len=x_emb.size(2)) # [B, N]
# encoder
enc = self.encoder(x_emb, token_mask=token_mask) # [B, D, N]
if use_d_gt is None:
# Default to False for all if not provided
use_d_gt = torch.zeros(x.size(0), dtype=torch.bool, device=x.device)
dur_int, dur_float = self.duration_predictor(enc, token_mask, min_dur=self.min_dur) # [B, N]
maskf = token_mask.float()
#dur_float = torch.clamp(dur_float, min=float(self.min_dur))
dur_float = dur_float * maskf # [B, N]
eps = 1e-6
maskf = token_mask.float() # [B, N]
d = dur_float * maskf # [B, N]
den = d.sum(dim=1, keepdim=True) # [B, 1]
# smooth proportions (stable even if den is tiny)
p = d / (den + eps) # [B, N]
Tb = frame_lengths.float().unsqueeze(1) # [B, 1]
l_keep = p * Tb # [B, N] scaled durations that sum ~ Tb
l_keep = l_keep * maskf
mse_raw = (dur_float - (gt_durations.float()* maskf)) ** 2
#mse_raw = (dur_float - (gt_durations.float() * maskf)).abs()
mse_raw = mse_raw * maskf
denom_mse_per = maskf.sum(dim=1)
L_mse_per = mse_raw.sum(dim=1) / denom_mse_per
L_mse = L_mse_per.mean()
dur_sum = (dur_float * maskf).sum(dim=1) # [B]
L_len = ((frame_lengths.float() - dur_sum) ** 2).mean()
upsampling_durations = torch.empty_like(dur_float)
for i in range(B):
if use_d_gt[i]:
upsampling_durations[i] = gt_durations[i].float()
else:
upsampling_durations[i] = dur_float[i]
frame_feats_keep, weights_keep = gaussian_upsample(
enc,
upsampling_durations,
frame_lengths,
sigma_sq=self.sigma_sq,
token_mask=token_mask,
)
info = {
"L_mse": L_mse,
"L_len": L_len,
"dur_float": dur_float,
}
return l_keep, info , frame_feats_keep , weights_keep
'''
if gt_durations is None:
raise ValueError("gt_durations must be provided when using GT-duration rewards.")
device = x.device
B = x.size(0)
# Prepare embeddings & token mask (compute token_mask after embedding so shapes are stable)
if self.use_input_is_embeds:
x_emb = x # [B, D, N]
else:
x_emb = self.embed(x).transpose(1, 2) # [B, D, N]
token_mask = length_to_mask_second(token_lengths, max_len=x_emb.size(2)) # [B, N]
# encoder
enc = self.encoder(x_emb, token_mask=token_mask) # [B, D, N]
if use_d_gt is None:
# Default to False for all if not provided
use_d_gt = torch.zeros(x.size(0), dtype=torch.bool, device=x.device)
dur_int, dur_float = self.duration_predictor(enc, token_mask, min_dur=self.min_dur) # [B, N]
maskf = token_mask.float()
#dur_float = torch.clamp(dur_float, min=float(self.min_dur))
dur_float = dur_float * maskf # [B, N]
dur_float_unlogged = (torch.exp(dur_float) - 1) * maskf
log_gt = torch.log(gt_durations.float() + 1) * maskf
mse_raw = (dur_float - log_gt) ** 2
#mse_raw = F.huber_loss(dur_float, log_gt, delta=0.4, reduction="none")
loss_lin_l1 = torch.abs(dur_float_unlogged - gt_durations)
pause_weight = 1.0 + (gt_durations / 33.0) * 2.0
# Final Combination
# We use a higher lambda (0.1) because your max_dur is small (33),
# so gradients won't explode like they would if max_dur was 1000.
loss_per_token = (mse_raw + (0.2 * loss_lin_l1 * pause_weight)) * maskf
mse_raw = mse_raw * maskf
#denom_mse = maskf.sum().clamp(min=1.0)
denom_mse_per = maskf.sum(dim=1).clamp(min=1.0) # [B]
# per-sample MSE: sum over tokens / valid_token_count
L_mse_per = loss_per_token.sum(dim=1) / denom_mse_per # [B]
L_mse = L_mse_per.mean()
#print(denom_mse_per)
#L_mse = mse_raw.sum() / denom_mse
dur_sum = (dur_float_unlogged * maskf).sum(dim=1) # [B]
L_len = ((frame_lengths.float() - dur_sum) ** 2).mean()
upsampling_durations = torch.empty_like(dur_float_unlogged)
for i in range(B):
if use_d_gt[i]:
upsampling_durations[i] = gt_durations[i].float()
else:
upsampling_durations[i] = dur_float_unlogged[i]
frame_feats_keep, weights_keep = gaussian_upsample(
enc,
upsampling_durations,
frame_lengths,
sigma_sq=self.sigma_sq,
token_mask=token_mask,
)
info = {
"L_mse": L_mse,
"L_len": L_len,
"dur_float": dur_float,
}
eps = 1e-6
maskf = token_mask.float() # [B, N]
d = dur_float_unlogged * maskf # [B, N]
den = d.sum(dim=1, keepdim=True) # [B, 1]
# smooth proportions (stable even if den is tiny)
p = d / (den + eps) # [B, N]
Tb = frame_lengths.float().unsqueeze(1) # [B, 1]
l_keep = p * Tb # [B, N] scaled durations that sum ~ Tb
l_keep = l_keep * maskf
return l_keep, info , frame_feats_keep , weights_keep
|