File size: 18,814 Bytes
251713e | 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 | """Stage 5: Two decoder heads on the shared latent z.
AsymmetricDecoder — z → pixel reconstruction (recon head)
UnderstandingDecoder — z → semantic vector aligned with vision teacher
(understanding head)
Both heads sit downstream of z so the latent must preserve enough information
for pixel-faithful reconstruction AND semantic alignment simultaneously.
UnifiedDetailExpander — cross-attends from target positions into z
PixelShuffleCNNDecoder — 4-stage PixelShuffle CNN (16× spatial upsample)
"""
from __future__ import annotations
from typing import Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
from mavt.model.transformer import StandardTransformerBlock
# --------------------------------------------------------------------------- #
# Position-encoded query generator #
# --------------------------------------------------------------------------- #
class FourDQueryEncoding(nn.Module):
"""Learnable 4D position encoding for decoder queries."""
def __init__(self, dim: int, max_t: int = 16, max_x: int = 64,
max_y: int = 64, max_z: int = 64):
super().__init__()
self.embed_t = nn.Embedding(max_t, dim // 4)
self.embed_x = nn.Embedding(max_x, dim // 4)
self.embed_y = nn.Embedding(max_y, dim // 4)
self.embed_z = nn.Embedding(max_z, dim // 4)
self.proj = nn.Linear(dim, dim)
def forward(self, positions: torch.Tensor, B: int) -> torch.Tensor:
"""positions: (N, 4) → queries (B, N, dim)"""
t = positions[:, 0].clamp(0, self.embed_t.num_embeddings - 1)
x = positions[:, 1].clamp(0, self.embed_x.num_embeddings - 1)
y = positions[:, 2].clamp(0, self.embed_y.num_embeddings - 1)
z = positions[:, 3].clamp(0, self.embed_z.num_embeddings - 1)
pe = torch.cat([self.embed_t(t), self.embed_x(x),
self.embed_y(y), self.embed_z(z)], dim=-1) # (N, dim)
pe = self.proj(pe).unsqueeze(0).expand(B, -1, -1) # (B, N, dim)
return pe
# --------------------------------------------------------------------------- #
# UnifiedDetailExpander #
# --------------------------------------------------------------------------- #
class UnifiedDetailExpander(nn.Module):
"""Inverts C-D Split: cross-attends from target grid to compressed z.
Uses 2 cross-attention layers from position-encoded queries into the
compressed VAE latent representation. When latent positions are supplied,
local residual detail tokens receive positional embeddings and a distance
bias so each output position prefers nearby detail while content stays
globally addressable.
"""
def __init__(self, latent_dim: int = 32, dec_dim: int = 768,
num_heads: int = 8, num_layers: int = 2,
local_detail_bias: float = 0.25):
super().__init__()
self.query_enc = FourDQueryEncoding(dec_dim)
self.kv_pos_enc = FourDQueryEncoding(latent_dim)
self.token_type_embed = nn.Embedding(2, latent_dim)
nn.init.zeros_(self.token_type_embed.weight)
self.kv_pos_scale = nn.Parameter(torch.tensor(0.1))
self.token_type_scale = nn.Parameter(torch.tensor(0.1))
self.norm_kv = nn.LayerNorm(latent_dim)
self.local_detail_bias = local_detail_bias
self.layers = nn.ModuleList([
nn.ModuleDict({
'norm_q': nn.LayerNorm(dec_dim),
'norm_ff': nn.LayerNorm(dec_dim),
'cross_attn': nn.MultiheadAttention(
embed_dim=dec_dim, num_heads=num_heads,
kdim=latent_dim, vdim=latent_dim,
batch_first=True, bias=True,
),
'ff': nn.Sequential(
nn.Linear(dec_dim, dec_dim * 4),
nn.GELU(),
nn.Linear(dec_dim * 4, dec_dim),
),
})
for _ in range(num_layers)
])
def _detail_distance_bias(
self,
target_positions: torch.Tensor,
latent_positions: Optional[torch.Tensor],
latent_token_types: Optional[torch.Tensor],
dtype: torch.dtype,
) -> Optional[torch.Tensor]:
if latent_positions is None or latent_token_types is None:
return None
detail_mask = latent_token_types == 1
if not bool(detail_mask.any()):
return None
q_pos = target_positions.float()
kv_pos = latent_positions.float()
dist = (q_pos[:, None, :] - kv_pos[None, :, :]).abs().sum(dim=-1)
bias = torch.zeros_like(dist, dtype=dtype)
bias[:, detail_mask] = -self.local_detail_bias * dist[:, detail_mask].to(dtype)
return bias
def forward(
self,
z: torch.Tensor,
target_positions: torch.Tensor,
latent_positions: Optional[torch.Tensor] = None,
latent_token_types: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""
z : (B, N_c+N_d, latent_dim)
target_positions: (N_target, 4)
Returns expanded : (B, N_target, dec_dim)
"""
B = z.shape[0]
q = self.query_enc(target_positions, B) # (B, N_target, dec_dim)
kv = z
if latent_positions is not None:
latent_positions = latent_positions.to(device=z.device, dtype=torch.long)
kv = kv + (
self.kv_pos_scale.to(kv.dtype)
* self.kv_pos_enc(latent_positions, B).to(kv.dtype)
)
if latent_token_types is not None:
latent_token_types = latent_token_types.to(device=z.device, dtype=torch.long)
kv = kv + (
self.token_type_scale.to(kv.dtype)
* self.token_type_embed(latent_token_types).unsqueeze(0).to(kv.dtype)
)
kv = self.norm_kv(kv) # (B, N_z, latent_dim)
attn_mask = self._detail_distance_bias(
target_positions.to(z.device),
latent_positions,
latent_token_types,
q.dtype,
)
for layer in self.layers:
q_n = layer['norm_q'](q)
out, _ = layer['cross_attn'](q_n, kv, kv, attn_mask=attn_mask)
q = q + out
q = q + layer['ff'](layer['norm_ff'](q))
return q # (B, N_target, dec_dim)
# --------------------------------------------------------------------------- #
# PixelShuffle CNN decoder #
# --------------------------------------------------------------------------- #
class ResBlock2D(nn.Module):
"""GroupNorm-GELU pre-activation residual block."""
def __init__(self, dim: int):
super().__init__()
groups = min(32, dim)
self.norm1 = nn.GroupNorm(groups, dim)
self.conv1 = nn.Conv2d(dim, dim, 3, padding=1)
self.norm2 = nn.GroupNorm(groups, dim)
self.conv2 = nn.Conv2d(dim, dim, 3, padding=1)
nn.init.zeros_(self.conv2.weight)
if self.conv2.bias is not None:
nn.init.zeros_(self.conv2.bias)
def forward(self, x: torch.Tensor) -> torch.Tensor:
h = self.conv1(F.gelu(self.norm1(x)))
h = self.conv2(F.gelu(self.norm2(h)))
return x + h
class WindowedSelfAttn2D(nn.Module):
"""Pre-LN windowed self-attention + MLP for 2D feature maps.
Partitions (B, C, H, W) into non-overlapping spatial windows of size
`window_size × window_size`. Attention runs INSIDE each window only, so
cost is O(B · nW · ws² · C) — feasible at 32²–64² where global attention
would be too heavy.
"""
def __init__(self, dim: int, num_heads: int = 8,
window_size: int = 8, mlp_ratio: float = 2.0):
super().__init__()
self.window_size = window_size
self.norm1 = nn.LayerNorm(dim)
self.attn = nn.MultiheadAttention(
embed_dim=dim, num_heads=num_heads,
batch_first=True, bias=True,
)
self.norm2 = nn.LayerNorm(dim)
mlp_dim = int(dim * mlp_ratio)
self.mlp = nn.Sequential(
nn.Linear(dim, mlp_dim),
nn.GELU(),
nn.Linear(mlp_dim, dim),
)
# Zero-init last MLP linear so block starts as identity (residual only).
nn.init.zeros_(self.mlp[-1].weight)
nn.init.zeros_(self.mlp[-1].bias)
def forward(self, x: torch.Tensor) -> torch.Tensor:
B, C, H, W = x.shape
ws = self.window_size
# Pad spatial dims if not divisible by window_size.
Hp = (H + ws - 1) // ws * ws
Wp = (W + ws - 1) // ws * ws
# Permute to channels-last for layernorm/attention.
xp = x.permute(0, 2, 3, 1) # (B, H, W, C)
if (Hp, Wp) != (H, W):
xp = F.pad(xp, (0, 0, 0, Wp - W, 0, Hp - H))
# Window partition: (B, nH, ws, nW, ws, C) → (B*nH*nW, ws*ws, C)
nH, nW = Hp // ws, Wp // ws
xp = xp.reshape(B, nH, ws, nW, ws, C)
xp = xp.permute(0, 1, 3, 2, 4, 5).reshape(-1, ws * ws, C)
# Pre-LN attention + MLP, both residual.
h = self.norm1(xp)
out, _ = self.attn(h, h, h)
xp = xp + out
xp = xp + self.mlp(self.norm2(xp))
# Reverse window partition.
xp = xp.reshape(B, nH, nW, ws, ws, C)
xp = xp.permute(0, 5, 1, 3, 2, 4).reshape(B, C, Hp, Wp)
return xp[..., :H, :W]
class PixelShuffleCNNDecoder(nn.Module):
"""4-stage progressive upsampler with CNN + windowed attention refinement.
Input : (B, in_channels, H_grid, W_grid) e.g. (B, 768, 16, 16)
Output: (B, 3, H_out, W_out) e.g. (B, 3, 256, 256), in [-1, 1]
Each stage upsamples 2× via Conv → PixelShuffle → GELU. Between stages we
insert ResBlock2D + WindowedSelfAttn2D so the model can refine features at
32², 64², 128² instead of leaving all high-freq generation to the final
Conv → PS layer alone (which previously had ~LPIPS bottleneck).
Param overhead vs flat CNN: ~+8 M (~30 M total).
"""
def __init__(self, in_channels: int = 768):
super().__init__()
# 16×16 → 32×32
self.up1 = nn.Sequential(
nn.Conv2d(in_channels, 512 * 4, 3, padding=1),
nn.PixelShuffle(2),
nn.GELU(),
)
self.refine1 = nn.Sequential(
ResBlock2D(512),
WindowedSelfAttn2D(512, num_heads=8, window_size=8),
)
# 32×32 → 64×64
self.up2 = nn.Sequential(
nn.Conv2d(512, 256 * 4, 3, padding=1),
nn.PixelShuffle(2),
nn.GELU(),
)
self.refine2 = nn.Sequential(
ResBlock2D(256),
WindowedSelfAttn2D(256, num_heads=8, window_size=8),
)
# 64×64 → 128×128 (no attn here — 16×16 windows = 256² tokens × 128 ch
# would dominate compute. ResBlock only.)
self.up3 = nn.Sequential(
nn.Conv2d(256, 128 * 4, 3, padding=1),
nn.PixelShuffle(2),
nn.GELU(),
)
self.refine3 = ResBlock2D(128)
# 128×128 → 256×256 (output, bounded to [-1, 1]).
self.up4 = nn.Sequential(
nn.Conv2d(128, 3 * 4, 3, padding=1),
nn.PixelShuffle(2),
nn.Tanh(),
)
self._icnr_init()
def _icnr_init(self) -> None:
"""ICNR (Aitken et al., 2017): initialise the r² sub-pixel filters of
each conv-before-PixelShuffle block identically so init acts like
nearest-neighbour upsample → no checkerboard artifact early on.
"""
r = 2 # PixelShuffle factor used in every block
for module in [self.up1, self.up2, self.up3, self.up4]:
for m in module:
if isinstance(m, nn.Conv2d) and m.out_channels % (r * r) == 0:
ni = m.in_channels
no = m.out_channels // (r * r)
kh, kw = m.kernel_size
kernel = m.weight.new_empty(no, ni, kh, kw)
nn.init.kaiming_normal_(kernel, nonlinearity='relu')
m.weight.data.copy_(kernel.repeat_interleave(r * r, dim=0))
if m.bias is not None:
nn.init.zeros_(m.bias)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.up1(x)
x = self.refine1(x)
x = self.up2(x)
x = self.refine2(x)
x = self.up3(x)
x = self.refine3(x)
x = self.up4(x)
return x
# --------------------------------------------------------------------------- #
# AsymmetricDecoder #
# --------------------------------------------------------------------------- #
class AsymmetricDecoder(nn.Module):
"""Full asymmetric decoder: expander → self-attention blocks → CNN upsample.
Handles all three modalities (image, video, 3D) with shared weights.
"""
def __init__(
self,
latent_dim: int = 32,
dec_dim: int = 768,
num_attn_blocks: int = 4,
num_heads: int = 12,
mlp_ratio: float = 4.0,
):
super().__init__()
self.expander = UnifiedDetailExpander(latent_dim, dec_dim, num_heads=num_heads)
self.self_attn_blocks = nn.ModuleList([
StandardTransformerBlock(dec_dim, num_heads, mlp_ratio)
for _ in range(num_attn_blocks)
])
self.cnn = PixelShuffleCNNDecoder(in_channels=dec_dim)
def _decode_grid(
self,
z: torch.Tensor, # (B, Nz, latent_dim)
positions: torch.Tensor, # (N_grid, 4)
H_grid: int,
W_grid: int,
latent_positions: Optional[torch.Tensor] = None,
latent_token_types: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""Decode a single 2D grid → (B, 3, H_out, W_out)."""
B = z.shape[0]
expanded = self.expander(
z, positions, latent_positions, latent_token_types
) # (B, N_grid, dec_dim)
for blk in self.self_attn_blocks:
expanded = blk(expanded)
# Reshape to 2D spatial grid
feat = expanded.transpose(1, 2).reshape(B, -1, H_grid, W_grid)
return self.cnn(feat) # (B, 3, 16*H_grid, 16*W_grid)
def forward(
self,
z: torch.Tensor, # (B, Nz, latent_dim)
target_positions: torch.Tensor, # (N_target, 4)
modality: str,
grid_shape: tuple, # (H_grid, W_grid) for image/3D; (Tp, H, W) for video
latent_positions: Optional[torch.Tensor] = None,
latent_token_types: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""Decode z → reconstructed pixel-space tensor."""
if modality == 'image':
H, W = grid_shape
out = self._decode_grid(
z, target_positions, H, W,
latent_positions=latent_positions,
latent_token_types=latent_token_types,
)
return out # (B, 3, H_out, W_out)
elif modality == 'video':
Tp, Hg, Wg = grid_shape
N_frame = Hg * Wg
# Decode per-frame with shared weights
frames = []
for t in range(Tp):
pos_t = target_positions[t * N_frame:(t + 1) * N_frame]
frame = self._decode_grid(
z, pos_t, Hg, Wg,
latent_positions=latent_positions,
latent_token_types=latent_token_types,
) # (B, 3, H, W)
frames.append(frame)
return torch.stack(frames, dim=2) # (B, 3, Tp, H, W)
elif modality == 'threed':
N_plane = target_positions.shape[0] // 3
Hg, Wg = grid_shape
planes_out = []
for p in range(3):
pos_p = target_positions[p * N_plane:(p + 1) * N_plane]
plane = self._decode_grid(
z, pos_p, Hg, Wg,
latent_positions=latent_positions,
latent_token_types=latent_token_types,
) # (B, 3, H, W)
planes_out.append(plane)
return torch.stack(planes_out, dim=1) # (B, 3, 3, H, W)
else:
raise ValueError(f"Unknown modality: {modality}")
# --------------------------------------------------------------------------- #
# UnderstandingDecoder #
# --------------------------------------------------------------------------- #
class UnderstandingDecoder(nn.Module):
"""Decode latent z → global semantic vector aligned with vision teacher.
Mirror of AsymmetricDecoder but for the understanding output. Operating on
z (the bottleneck) — not on the encoder's pre-VAE features — forces the
latent to preserve enough semantic information to recover a SigLIP-aligned
representation. This is the "understanding head" of the unified tokenizer.
Architecture:
z (B, Nz, latent_dim)
→ Linear(latent_dim → dec_dim) + LayerNorm
→ N self-attention blocks (refine token interactions)
→ attention pool with single learnable query → (B, dec_dim)
→ LayerNorm + Linear(dec_dim → semantic_dim)
"""
def __init__(
self,
latent_dim: int = 32,
dec_dim: int = 768,
semantic_dim: int = 768,
num_heads: int = 8,
num_layers: int = 2,
mlp_ratio: float = 4.0,
):
super().__init__()
self.in_proj = nn.Linear(latent_dim, dec_dim)
self.norm_in = nn.LayerNorm(dec_dim)
self.self_attn_blocks = nn.ModuleList([
StandardTransformerBlock(dec_dim, num_heads, mlp_ratio)
for _ in range(num_layers)
])
self.query = nn.Parameter(torch.randn(1, 1, dec_dim) * (dec_dim ** -0.5))
self.pool = nn.MultiheadAttention(
embed_dim=dec_dim, num_heads=num_heads,
batch_first=True, bias=True,
)
self.norm_out = nn.LayerNorm(dec_dim)
self.proj = nn.Linear(dec_dim, semantic_dim)
def forward(self, z: torch.Tensor) -> torch.Tensor:
"""z: (B, Nz, latent_dim) → semantic: (B, semantic_dim)"""
x = self.norm_in(self.in_proj(z))
for blk in self.self_attn_blocks:
x = blk(x)
B = x.shape[0]
q = self.query.expand(B, 1, -1)
pooled, _ = self.pool(q, x, x)
return self.proj(self.norm_out(pooled.squeeze(1)))
|