File size: 17,369 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 | import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange
from torch import Tensor
from math import log, pi
from typing import Callable, Optional, Tuple
# -----------------------------------------------------------------------------
# Helper functions
# -----------------------------------------------------------------------------
def exists(x) -> bool:
return x is not None
def default(val, fn):
return val if val is not None else fn()
def rand_bool(shape, proba: float, device: torch.device) -> Tensor:
"""Sample Bernoulli mask with given probability of True."""
return torch.rand(shape, device=device) < proba
# -----------------------------------------------------------------------------
# Normalization / Transformer components
# -----------------------------------------------------------------------------
class RMSNorm(nn.Module):
"""RMSNorm (does not subtract mean)."""
def __init__(
self,
dim: int,
eps: float = 1e-8,
elementwise_affine: bool = True,
):
super().__init__()
self.dim = dim
self.eps = eps
self.elementwise_affine = elementwise_affine
if elementwise_affine:
self.weight = nn.Parameter(torch.ones(dim))
else:
self.register_buffer("weight", torch.ones(dim))
def forward(self, x: Tensor) -> Tensor:
# x: (..., dim)
rms = x.pow(2).mean(dim=-1, keepdim=True).add(self.eps).sqrt()
x_normed = x / rms
return x_normed * self.weight
def _make_norm(norm_type: str, dim: int) -> nn.Module:
"""Create a norm module by name."""
if norm_type is None or norm_type == "none":
return nn.Identity()
if norm_type == "layer":
return nn.LayerNorm(dim)
if norm_type == "rms":
return RMSNorm(dim)
raise ValueError(f"Unknown norm_type: {norm_type}")
class RotaryEmbedding(nn.Module):
"""
RoPE implementation that generates sin/cos on demand.
Works on head dimension d (must be even).
"""
def __init__(self, dim: int, max_seq_len: int = 512, base: int = 10000):
super().__init__()
assert dim % 2 == 0, "Rotary embedding dim must be even"
self.dim = dim
self.max_seq_len = max_seq_len
self.base = base
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
self.register_buffer("inv_freq", inv_freq, persistent=False)
def _build_sin_cos(
self,
seq_len: int,
device: torch.device,
dtype: torch.dtype,
) -> Tuple[Tensor, Tensor]:
positions = torch.arange(seq_len, device=device, dtype=dtype).unsqueeze(1)
angles = positions * rearrange(
self.inv_freq.to(device=device, dtype=dtype),
"d -> 1 d",
)
sin = torch.sin(angles)
cos = torch.cos(angles)
sin = torch.stack([sin, sin], dim=-1).reshape(seq_len, self.dim)
cos = torch.stack([cos, cos], dim=-1).reshape(seq_len, self.dim)
return sin, cos
@staticmethod
def rotate_half(x: Tensor) -> Tensor:
x1 = x[..., ::2]
x2 = x[..., 1::2]
x_rotated = torch.stack((-x2, x1), dim=-1).reshape_as(x)
return x_rotated
def apply_rotary(self, q: Tensor, k: Tensor) -> Tuple[Tensor, Tensor]:
"""
q, k: (b, h, n, d) with d == self.dim
"""
assert q.shape[-1] == self.dim and k.shape[-1] == self.dim
seq_len = q.shape[-2]
device = q.device
dtype = q.dtype
sin, cos = self._build_sin_cos(seq_len, device=device, dtype=dtype)
sin = sin.unsqueeze(0).unsqueeze(0) # [1,1,n,d]
cos = cos.unsqueeze(0).unsqueeze(0)
q_out = (q * cos) + (self.rotate_half(q) * sin)
k_out = (k * cos) + (self.rotate_half(k) * sin)
return q_out, k_out
def FeedForward(features: int, multiplier: int) -> nn.Module:
mid_features = features * multiplier
return nn.Sequential(
nn.Linear(features, mid_features),
nn.GELU(),
nn.Linear(mid_features, features),
)
class AttentionBase(nn.Module):
def __init__(
self,
features: int,
*,
head_features: int,
num_heads: int,
use_rope: bool,
rope_max_seq_len: int = 512,
out_features: Optional[int] = None,
attn_dropout: float = 0.0,
out_dropout: float = 0.0,
):
super().__init__()
self.scale = head_features ** -0.5
self.num_heads = num_heads
self.use_rope = use_rope
mid_features = head_features * num_heads
if out_features is None:
out_features = features
self.to_out = nn.Linear(mid_features, out_features)
self.attn_dropout = (
nn.Dropout(attn_dropout) if attn_dropout > 0.0 else nn.Identity()
)
self.out_dropout = (
nn.Dropout(out_dropout) if out_dropout > 0.0 else nn.Identity()
)
if use_rope:
self.rotary = RotaryEmbedding(
dim=head_features,
max_seq_len=rope_max_seq_len,
)
else:
self.rotary = None
def forward(self, q: Tensor, k: Tensor, v: Tensor, mask: Optional[Tensor] = None) -> Tensor:
# q,k,v: (b, n, h*d)
q = rearrange(q, "b n (h d) -> b h n d", h=self.num_heads)
k = rearrange(k, "b n (h d) -> b h n d", h=self.num_heads)
v = rearrange(v, "b n (h d) -> b h n d", h=self.num_heads)
if self.rotary is not None:
q, k = self.rotary.apply_rotary(q, k)
# sim: [Batch, Heads, QueryLen, KeyLen]
sim = torch.einsum("... n d, ... m d -> ... n m", q, k)
sim = sim * self.scale
# --- MASKING APPLIED HERE ---
if mask is not None:
# mask should be broadcastable to [Batch, Heads, QueryLen, KeyLen]
# Usually mask is [B, 1, 1, KeyLen] where True=Keep, False=Mask
# We use a large negative number instead of -inf to be safe with fp16 mixed precision
mask_value = -torch.finfo(sim.dtype).max if sim.dtype == torch.float32 else -1e4
sim = sim.masked_fill(~mask, mask_value)
# ----------------------------
attn = sim.softmax(dim=-1)
attn = self.attn_dropout(attn)
out = torch.einsum("... n m, ... m d -> ... n d", attn, v)
out = rearrange(out, "b h n d -> b n (h d)")
out = self.to_out(out)
out = self.out_dropout(out)
return out
class Attention(nn.Module):
def __init__(
self,
features: int,
*,
head_features: int,
num_heads: int,
out_features: Optional[int] = None,
use_rope: bool,
rope_max_seq_len: int = 512,
attn_dropout: float = 0.0,
out_dropout: float = 0.0,
norm_type: str = "layer",
):
super().__init__()
mid_features = head_features * num_heads
self.norm = _make_norm(norm_type, features)
self.to_q = nn.Linear(features, mid_features, bias=False)
self.to_kv = nn.Linear(features, mid_features * 2, bias=False)
self.attention = AttentionBase(
features=features,
out_features=out_features,
num_heads=num_heads,
head_features=head_features,
use_rope=use_rope,
rope_max_seq_len=rope_max_seq_len,
attn_dropout=attn_dropout,
out_dropout=out_dropout,
)
def forward(self, x: Tensor, mask: Optional[Tensor] = None) -> Tensor:
if not isinstance(self.norm, nn.Identity):
x_norm = self.norm(x)
else:
x_norm = x
q = self.to_q(x_norm)
kv = self.to_kv(x_norm)
k, v = torch.chunk(kv, 2, dim=-1)
# Pass mask down
return self.attention(q, k, v, mask=mask)
class TransformerBlock(nn.Module):
def __init__(
self,
features: int,
num_heads: int,
head_features: int,
multiplier: int,
use_rope: bool,
rope_max_seq_len: int = 512,
dropout: float = 0.0,
attn_dropout: float = 0.0,
ff_dropout: float = 0.0,
norm_type: str = "layer",
):
super().__init__()
self.attention = Attention(
features=features,
num_heads=num_heads,
head_features=head_features,
use_rope=use_rope,
rope_max_seq_len=rope_max_seq_len,
attn_dropout=attn_dropout,
out_dropout=dropout,
norm_type=norm_type,
)
self.feed_forward = FeedForward(features=features, multiplier=multiplier)
self.norm_ff = _make_norm(norm_type, features)
self.ff_dropout = (
nn.Dropout(ff_dropout) if ff_dropout > 0.0 else nn.Identity()
)
def forward(self, x: Tensor, mask: Optional[Tensor] = None) -> Tensor:
# Pass mask to attention
x = self.attention(x, mask=mask) + x
if not isinstance(self.norm_ff, nn.Identity):
ff_in = self.norm_ff(x)
else:
ff_in = x
ff_out = self.feed_forward(ff_in)
ff_out = self.ff_dropout(ff_out)
x = ff_out + x
return x
# -----------------------------------------------------------------------------
# Time and fixed positional embeddings
# -----------------------------------------------------------------------------
class LearnedPositionalEmbedding(nn.Module):
"""Continuous time embedding."""
def __init__(self, dim: int):
super().__init__()
assert dim % 2 == 0
half_dim = dim // 2
self.weights = nn.Parameter(torch.randn(half_dim))
def forward(self, x: Tensor) -> Tensor:
x = rearrange(x, "b -> b 1")
freqs = x * rearrange(self.weights, "d -> 1 d") * 2 * pi
fouriered = torch.cat((freqs.sin(), freqs.cos()), dim=-1)
fouriered = torch.cat((x, fouriered), dim=-1)
return fouriered
def TimePositionalEmbedding(dim: int, out_features: int) -> nn.Module:
return nn.Sequential(
LearnedPositionalEmbedding(dim),
nn.Linear(in_features=dim + 1, out_features=out_features),
)
class FixedEmbedding(nn.Module):
def __init__(self, max_length: int, features: int):
super().__init__()
self.max_length = max_length
self.embedding = nn.Embedding(max_length, features)
def forward(self, x: Tensor) -> Tensor:
"""
x: [B, T, D] (only T and B are used, D is ignored)
returns: [B, T, features]
"""
batch_size, length = x.shape[0], x.shape[1]
device = x.device
assert (
length <= self.max_length
), "Input sequence length must be <= max_length"
position = torch.arange(length, device=device)
fixed_embedding = self.embedding(position) # [T, features]
fixed_embedding = fixed_embedding.unsqueeze(0).expand(
batch_size,
length,
-1,
)
return fixed_embedding
# -----------------------------------------------------------------------------
# Seq-to-seq diffusion net for joint pitch + energy
# -----------------------------------------------------------------------------
class PitchEnergyTransformer1d(nn.Module):
"""
Seq-to-seq transformer used as diffusion network.
Input:
x : [B, 2, T] (2 channels: pitch, energy)
time : [B] (noise level embedding, from VKDiffusion)
embedding : [B, 512, T] (per-frame conditioning, e.g. P from predictor)
mask : [B, T] (Boolean padding mask, True=Keep, False=Pad)
Output:
v_pred : [B, 2, T]
"""
def __init__(
self,
*,
num_layers: int = 6,
num_heads: int = 8,
head_features: int = 64,
multiplier: int = 4,
ctx_dim: int = 512,
max_seq_len: int = 2048,
use_rope: bool = True,
dropout: float = 0.0,
attn_dropout: float = 0.0,
ff_dropout: float = 0.0,
norm_type: str = "rms",
):
super().__init__()
self.data_channels = 2
self.ctx_dim = ctx_dim
self.max_seq_len = max_seq_len
total_features = self.data_channels + ctx_dim
self.input_dropout = (
nn.Dropout(dropout) if dropout > 0.0 else nn.Identity()
)
self.blocks = nn.ModuleList(
[
TransformerBlock(
features=total_features,
num_heads=num_heads,
head_features=head_features,
multiplier=multiplier,
use_rope=use_rope,
rope_max_seq_len=max_seq_len,
dropout=dropout,
attn_dropout=attn_dropout,
ff_dropout=ff_dropout,
norm_type=norm_type,
)
for _ in range(num_layers)
]
)
self.to_out = nn.Linear(total_features, self.data_channels)
# time embedding
self.to_time = nn.Sequential(
TimePositionalEmbedding(
dim=self.data_channels,
out_features=total_features,
),
nn.GELU(),
)
self.to_mapping = nn.Sequential(
nn.Linear(total_features, total_features),
nn.GELU(),
nn.Linear(total_features, total_features),
nn.GELU(),
)
self.fixed_embedding = FixedEmbedding(
max_length=max_seq_len,
features=ctx_dim,
)
def get_mapping(self, time: Tensor) -> Tensor:
mapping = self.to_time(time)
mapping = self.to_mapping(mapping)
return mapping # [B, F]
def _run_core(
self,
x: Tensor,
time: Tensor,
ctx: Tensor,
mask: Optional[Tensor] = None
) -> Tensor:
"""
x : [B, 2, T]
ctx : [B, ctx_dim, T]
mask: [B, T] (True=Keep, False=Pad)
"""
b, c, t = x.shape
assert c == self.data_channels
assert (
ctx.shape[0] == b
and ctx.shape[1] == self.ctx_dim
and ctx.shape[2] == t
)
x_seq = x.transpose(1, 2) # [B, T, 2]
ctx_seq = ctx.transpose(1, 2) # [B, T, ctx_dim]
h = torch.cat([x_seq, ctx_seq], dim=-1) # [B, T, F]
mapping = self.get_mapping(time).unsqueeze(1) # [B,1,F]
h = self.input_dropout(h)
# --- PREPARE ATTENTION MASK ---
# If mask is present, we reshape it for broadcasting in AttentionBase
# Input mask: [B, T]
# Target shape for attention: [B, 1, 1, T] (Broadcasting over Heads and QueryLen)
attn_mask = None
if mask is not None:
attn_mask = mask.unsqueeze(1).unsqueeze(1) # [B, 1, 1, T]
# ------------------------------
for block in self.blocks:
h = block(h + mapping, mask=attn_mask)
h = self.to_out(h) # [B, T, 2]
out = h.transpose(1, 2) # [B, 2, T]
return out
def forward(
self,
x: Tensor,
time: Tensor,
*,
mask: Optional[Tensor] = None,
embedding: Tensor,
embedding_mask_proba: float = 0.0,
embedding_scale: float = 1.0,
) -> Tensor:
"""
x : [B, 2, T]
time : [B]
embedding : [B, 512, T]
mask : [B, T] (Padding mask. True=Valid Frame, False=Pad)
"""
assert embedding is not None
b, ctx_dim, t = embedding.shape
assert ctx_dim == self.ctx_dim
device = embedding.device
emb_seq = embedding.transpose(1, 2) # [B, T, ctx_dim]
fixed_emb_seq = self.fixed_embedding(emb_seq) # [B, T, ctx_dim]
# Apply dropout to the *conditioning embedding* (Classifier-Free Guidance style)
if embedding_mask_proba > 0.0:
rand_mask = rand_bool(
shape=(b, 1, 1),
proba=embedding_mask_proba,
device=device,
)
emb_seq = torch.where(rand_mask, fixed_emb_seq, emb_seq)
emb_ctx = emb_seq.transpose(1, 2) # [B, ctx_dim, T]
fixed_ctx = fixed_emb_seq.transpose(1, 2) # [B, ctx_dim, T]
if embedding_scale != 1.0:
# CFG: We must pass the padding mask to both calls
out_cond = self._run_core(x, time, ctx=emb_ctx, mask=mask)
out_uncond = self._run_core(x, time, ctx=fixed_ctx, mask=mask)
return out_uncond + (out_cond - out_uncond) * embedding_scale
return self._run_core(x, time, ctx=emb_ctx, mask=mask) |