File size: 18,592 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 | import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional, Tuple
from einops import repeat
from einops.layers.torch import Rearrange
# -----------------------------------------------------------------------------
# small utils (keep yours if already present)
# -----------------------------------------------------------------------------
def exists(val):
return val is not None
def rand_bool(shape, proba, device=None):
if proba == 1:
return torch.ones(shape, device=device, dtype=torch.bool)
if proba == 0:
return torch.zeros(shape, device=device, dtype=torch.bool)
return torch.bernoulli(torch.full(shape, proba, device=device)).to(torch.bool)
# -----------------------------------------------------------------------------
# time embedding (reuse your existing ones if already in this file)
# -----------------------------------------------------------------------------
class LearnedPositionalEmbedding(nn.Module):
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: torch.Tensor) -> torch.Tensor:
# x: [B]
x = x.view(-1, 1)
freqs = x * self.weights.view(1, -1) * 2 * math.pi
fouriered = torch.cat((freqs.sin(), freqs.cos()), dim=-1)
return torch.cat((x, fouriered), dim=-1)
def TimePositionalEmbedding(dim: int, out_features: int) -> nn.Module:
return nn.Sequential(
LearnedPositionalEmbedding(dim),
nn.Linear(dim + 1, out_features),
)
class FixedEmbedding(nn.Module):
"""
Safer than your current version: supports any length by wrapping positions.
"""
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: torch.Tensor) -> torch.Tensor:
b, t = x.shape[0], x.shape[1]
device = x.device
pos = torch.arange(t, device=device) % self.max_length
emb = self.embedding(pos) # [T, D]
return repeat(emb, "t d -> b t d", b=b)
# -----------------------------------------------------------------------------
# HRM building blocks (RoPE attention + SwiGLU FFN + AdaRMSNorm)
# -----------------------------------------------------------------------------
class AdaRMSNorm(nn.Module):
def __init__(self, style_dim: int, channels: int, eps: float = 1e-5):
super().__init__()
self.fc = nn.Linear(style_dim, channels * 2)
self.eps = eps
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)
gamma, beta = gb.chunk(2, dim=-1)
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):
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
class RotaryEmbedding(nn.Module):
def __init__(self, dim: int, base: float = 10000.0):
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) # [T, head_dim]
cos = emb.cos().unsqueeze(0).unsqueeze(0) # [1, 1, T, D]
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: [B, H, T, D]
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,
):
super().__init__()
if embed_dim % num_heads != 0:
raise ValueError("embed_dim must be divisible by num_heads.")
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
if use_rope and (self.head_dim % 2 != 0):
raise ValueError(
f"RoPE requires even head_dim, got head_dim={self.head_dim}."
)
def forward(
self,
x: torch.Tensor,
key_padding_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
# x: [B, T, C], key_padding_mask: [B, T] with True = PAD
b, t, c = x.shape
q, k, v = self.qkv(x).chunk(3, dim=-1)
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:
pad = key_padding_mask.unsqueeze(1).unsqueeze(2) # [B,1,1,T]
attn_scores = attn_scores.masked_fill(
pad, torch.finfo(attn_scores.dtype).min
)
attn = F.softmax(attn_scores, dim=-1)
attn = self.attn_dropout(attn)
out = torch.matmul(attn, v) # [B,H,T,D]
out = out.transpose(1, 2).contiguous().view(b, t, c)
return self.out_proj(out)
class HRMReasoningBlock(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,
use_rope: bool = True,
rope_base: float = 10000.0,
rms_eps: float = 1e-5,
):
super().__init__()
self.attn = MultiheadSelfAttentionRoPE(
dim, num_heads, attn_dropout, use_rope, rope_base
)
self.ff = SwiGLUFFN(dim, expansion, resid_dropout)
self.resid_dropout = nn.Dropout(resid_dropout)
self.ada1 = AdaRMSNorm(style_dim, dim, eps=rms_eps)
self.ada2 = AdaRMSNorm(style_dim, dim, eps=rms_eps)
def forward(
self,
hidden_states: torch.Tensor,
style: torch.Tensor,
key_padding_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
attn_out = self.attn(hidden_states, key_padding_mask=key_padding_mask)
hidden_states = self.ada1(
hidden_states + self.resid_dropout(attn_out), style
)
mlp_out = self.ff(hidden_states)
hidden_states = self.ada2(
hidden_states + self.resid_dropout(mlp_out), style
)
return hidden_states
class HRMReasoningModule(nn.Module):
def __init__(self, layers: nn.ModuleList):
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] = None,
) -> torch.Tensor:
hidden_states = hidden_states + input_injection
for layer in self.layers:
hidden_states = layer(
hidden_states, style, key_padding_mask=key_padding_mask
)
return hidden_states
# -----------------------------------------------------------------------------
# HRM-based StyleTransformer1d (same external API as your current one)
# -----------------------------------------------------------------------------
class StyleTransformer1d(nn.Module):
"""
HRM replacement for your StyleTransformer1d.
- Works on tokens: concat([x, embedding]) => model_dim = channels + cond_dim
- Uses HRM H/L cycles instead of stacked transformer blocks
- Style-conditioning via AdaRMSNorm in every block
- Time/features mapping added as a bias/injection
"""
def __init__(
self,
num_layers: int,
channels: int,
num_heads: int,
head_features: int, # kept for signature compatibility; ignored
multiplier: int,
use_context_time: bool = True,
use_rel_pos: bool = False, # ignored
context_features_multiplier: int = 1, # ignored
rel_pos_num_buckets: Optional[int] = None, # ignored
rel_pos_max_distance: Optional[int] = None, # ignored
context_features: Optional[int] = None, # style dim
context_embedding_features: Optional[int] = None, # cond dim
embedding_max_length: int = 512,
H_cycles: int = 1,
L_cycles: int = 2,
attn_dropout: float = 0.0,
resid_dropout: float = 0.0,
use_rope: bool = True,
rope_base: float = 10000.0,
):
super().__init__()
assert exists(context_embedding_features), "context_embedding_features is required"
assert exists(context_features), "context_features (style dim) is required"
self.channels = channels
self.cond_dim = context_embedding_features
self.style_dim = context_features
self.model_dim = channels + context_embedding_features
self.use_context_time = use_context_time
self.use_context_features = exists(context_features)
self.H_cycles = H_cycles
self.L_cycles = L_cycles
# Optional positional embedding added to conditioning sequence
self.fixed_embedding = FixedEmbedding(
max_length=embedding_max_length,
features=context_embedding_features,
)
# time/features mapping -> model_dim (same as concat(x, embedding))
if use_context_time or self.use_context_features:
self.to_mapping = nn.Sequential(
nn.Linear(self.model_dim, self.model_dim),
nn.GELU(),
nn.Linear(self.model_dim, self.model_dim),
nn.GELU(),
)
if use_context_time:
self.to_time = nn.Sequential(
TimePositionalEmbedding(dim=channels, out_features=self.model_dim),
nn.GELU(),
)
if self.use_context_features:
self.to_features = nn.Sequential(
nn.Linear(context_features, self.model_dim),
nn.GELU(),
)
# HRM states + step embeddings
self.H_init = nn.Parameter(torch.zeros(1, 1, self.model_dim))
self.L_init = nn.Parameter(torch.zeros(1, 1, self.model_dim))
nn.init.trunc_normal_(self.H_init, std=0.02)
nn.init.trunc_normal_(self.L_init, std=0.02)
self.h_step_emb = nn.Embedding(H_cycles, self.model_dim)
self.l_step_emb = nn.Embedding(L_cycles, self.model_dim)
layers_L = nn.ModuleList(
[
HRMReasoningBlock(
dim=self.model_dim,
num_heads=num_heads,
expansion=float(multiplier),
style_dim=self.style_dim,
attn_dropout=attn_dropout,
resid_dropout=resid_dropout,
use_rope=use_rope,
rope_base=rope_base,
)
for _ in range(num_layers)
]
)
layers_H = nn.ModuleList(
[
HRMReasoningBlock(
dim=self.model_dim,
num_heads=num_heads,
expansion=float(multiplier),
style_dim=self.style_dim,
attn_dropout=attn_dropout,
resid_dropout=resid_dropout,
use_rope=use_rope,
rope_base=rope_base,
)
for _ in range(num_layers)
]
)
self.L_level = HRMReasoningModule(layers_L)
self.H_level = HRMReasoningModule(layers_H)
# project back to `channels`
self.to_out = nn.Sequential(
Rearrange("b t c -> b c t"),
nn.Conv1d(self.model_dim, channels, kernel_size=1),
)
def get_mapping(
self,
time: Optional[torch.Tensor] = None,
features: Optional[torch.Tensor] = None,
) -> Optional[torch.Tensor]:
items = []
if self.use_context_time:
assert exists(time), "use_context_time=True but time was not provided"
items.append(self.to_time(time))
if self.use_context_features:
assert exists(features), "context_features exists but features were not provided"
items.append(self.to_features(features))
if len(items) == 0:
return None
mapping = torch.stack(items, dim=0).sum(dim=0) # [B, model_dim]
return self.to_mapping(mapping)
def run(
self,
x: torch.Tensor, # [B, T, channels]
time: torch.Tensor, # [B]a
embedding: torch.Tensor, # [B, T, cond_dim]
features: torch.Tensor, # [B, style_dim]
key_padding_mask: Optional[torch.Tensor] = None, # [B, T] True=PAD
) -> torch.Tensor:
# add positional embedding to cond (this was computed but unused in your old code)
embedding = embedding + self.fixed_embedding(embedding).to(dtype=embedding.dtype)
# concat token-wise
x_cat = torch.cat([x, embedding], dim=-1) # [B, T, model_dim]
mapping = self.get_mapping(time=time, features=features)
if mapping is not None:
x_cat = x_cat + mapping.unsqueeze(1)
b, t, _ = x_cat.shape
z_H = self.H_init.to(dtype=x_cat.dtype).expand(b, t, -1)
z_L = self.L_init.to(dtype=x_cat.dtype).expand(b, t, -1)
for h in range(self.H_cycles):
h_emb = self.h_step_emb.weight[h].view(1, 1, -1).to(dtype=x_cat.dtype)
z_H_curr = z_H + h_emb
for l in range(self.L_cycles):
l_emb = self.l_step_emb.weight[l].view(1, 1, -1).to(dtype=x_cat.dtype)
inj = z_H_curr + x_cat + l_emb
z_L = self.L_level(
z_L,
input_injection=inj,
style=features,
key_padding_mask=key_padding_mask,
)
if key_padding_mask is not None:
z_L = z_L.masked_fill(key_padding_mask.unsqueeze(-1), 0.0)
z_H = self.H_level(
z_H,
input_injection=z_L,
style=features,
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)
out = self.to_out(z_H).transpose(1, 2) # [B, T, channels]
return out
def forward(
self,
x: torch.Tensor,
time: torch.Tensor,
embedding_mask_proba: float = 0.0,
embedding: Optional[torch.Tensor] = None,
features: Optional[torch.Tensor] = None,
embedding_scale: float = 1.0,
key_padding_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
assert exists(embedding), "embedding is required"
assert exists(features), "features(style) is required"
b = embedding.shape[0]
device = embedding.device
if embedding_mask_proba > 0.0:
# Shape [B, 1, 1] for broadcasting
batch_mask = rand_bool((b, 1, 1), embedding_mask_proba, device=device)
# Mask the embedding [B, T, D]
embedding = torch.where(batch_mask, torch.zeros_like(embedding), embedding)
# Mask the features [B, S] -> squeeze mask to [B, 1] for broadcasting
features_mask = batch_mask.squeeze(-1) # [B, 1]
features = torch.where(features_mask, torch.zeros_like(features), features)
if embedding_scale != 1.0:
out = self.run(
x=x,
time=time,
embedding=embedding,
features=features,
key_padding_mask=key_padding_mask,
)
null_embedding = torch.zeros_like(embedding)
out_null = self.run(
x=x,
time=time,
embedding=null_embedding,
features=features,
key_padding_mask=key_padding_mask,
)
return out_null + (out - out_null) * embedding_scale
return self.run(
x=x,
time=time,
embedding=embedding,
features=features,
key_padding_mask=key_padding_mask,
) |