File size: 11,680 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 | """Stage 3: Content-Detail Split via slot cross-attention.
ContentExtractor → N_c content tokens (0.25·N by default)
LocalDetailPooler → local residual detail tokens
Monitoring signals (logged during training):
slot_diversity : mean pairwise cosine sim of content slots (target ≤ 0.5)
residual_ratio : ||R|| / ||x|| (target 0.3–0.5)
detail_contribution : variance fraction from detail branch
"""
from __future__ import annotations
from typing import Dict, Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
class CrossAttentionLayer(nn.Module):
"""Single cross-attention + FFN layer (pre-LN)."""
def __init__(self, dim: int, num_heads: int = 8, kv_dim: Optional[int] = None,
mlp_ratio: float = 4.0):
super().__init__()
kv_dim = kv_dim or dim
self.norm_q = nn.LayerNorm(dim)
self.norm_kv = nn.LayerNorm(kv_dim)
self.norm_ff = nn.LayerNorm(dim)
self.attn = nn.MultiheadAttention(
embed_dim=dim, num_heads=num_heads,
kdim=kv_dim, vdim=kv_dim,
batch_first=True, bias=True,
)
mlp_dim = int(dim * mlp_ratio)
self.ff = nn.Sequential(
nn.Linear(dim, mlp_dim),
nn.GELU(),
nn.Linear(mlp_dim, dim),
)
def forward(self, q: torch.Tensor, kv: torch.Tensor) -> torch.Tensor:
# q: (B, Nq, D), kv: (B, Nkv, D_kv)
q = self.norm_q(q)
k = self.norm_kv(kv)
out, _ = self.attn(q, k, k)
q = q + out
q = q + self.ff(self.norm_ff(q))
return q
class SlotPooler(nn.Module):
"""Slot cross-attention pooler: learns to pool N tokens into num_slots tokens."""
def __init__(self, num_slots: int, dim: int, num_heads: int = 8,
num_layers: int = 2):
super().__init__()
self.num_slots = num_slots
# Learnable slot initialisation
self.slots = nn.Parameter(torch.randn(1, num_slots, dim) * (dim ** -0.5))
self.layers = nn.ModuleList([
CrossAttentionLayer(dim, num_heads) for _ in range(num_layers)
])
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""x: (B, N, D) → slots: (B, num_slots, D)"""
B = x.shape[0]
slots = self.slots.expand(B, -1, -1)
for layer in self.layers:
slots = layer(slots, x)
return slots
class ContentDetailSplit(nn.Module):
"""Content-Detail Split module.
Separates tokens into a content channel (semantic, low-frequency) and a
detail channel (residual, high-frequency).
Content stays global: learned slot attention pools the full token sequence
into semantic / low-frequency slots. Detail is local: residual tokens are
pooled inside small coordinate windows, preserving a window-center position
for each detail token. The decoder can then prefer nearby detail tokens
instead of reconstructing texture from positionless global slots.
Note on parameter registration:
Content slot poolers depend on N_c which depends on modality / resolution.
Call ``prepare_poolers(N_c, N_d)`` for every combo that will appear at
training time BEFORE the optimizer is built — otherwise the pooler
params are not in any param_group and never receive updates. The lazy
fallback in ``_get_content_pooler`` only exists to keep smoke tests and
one-off inference paths functional; it emits a ``RuntimeWarning``.
"""
def __init__(
self,
dim: int = 768,
num_heads: int = 8,
num_slot_layers: int = 2,
local_detail_window_size: int = 1,
local_detail_temporal_window_size: int = 1,
):
super().__init__()
self.dim = dim
self.local_detail_window_size = local_detail_window_size
self.local_detail_temporal_window_size = local_detail_temporal_window_size
# Content slots are built dynamically based on (N, content_ratio);
# the key keeps N_d for backward-compatible checkpoint naming.
self._content_poolers: nn.ModuleDict = nn.ModuleDict()
self._num_heads = num_heads
self._num_slot_layers = num_slot_layers
self.detail_norm = nn.LayerNorm(dim)
self.detail_proj = nn.Linear(dim, dim)
def prepare_poolers(self, N_c: int, N_d: int) -> None:
"""Eagerly create content poolers for a known (N_c, N_d) combo.
Call once per expected combo BEFORE ``configure_optimizers`` runs so
that the new params are picked up by the optimizer's param_groups.
"""
key = f"{N_c}_{N_d}"
if key in self._content_poolers:
return
self._content_poolers[key] = SlotPooler(
N_c, self.dim, self._num_heads, self._num_slot_layers)
def _get_content_pooler(self, N_c: int, N_d: int) -> SlotPooler:
key = f"{N_c}_{N_d}"
if key not in self._content_poolers:
import warnings
warnings.warn(
f"ContentDetailSplit: lazy pooler creation for "
f"(N_c={N_c}, N_d={N_d}); its params are NOT in the "
f"optimizer and will stay at random init. Call "
f"prepare_poolers() in setup() before configure_optimizers().",
RuntimeWarning,
stacklevel=2,
)
self.prepare_poolers(N_c, N_d)
return self._content_poolers[key]
@staticmethod
def _default_positions(N: int, device: torch.device) -> torch.Tensor:
"""Fallback positions for direct unit tests without patch metadata."""
side = int(N ** 0.5)
pos = torch.zeros(N, 4, dtype=torch.long, device=device)
if side * side == N:
i = torch.arange(side, device=device)
j = torch.arange(side, device=device)
gi, gj = torch.meshgrid(i, j, indexing='ij')
pos[:, 1] = gi.reshape(-1)
pos[:, 2] = gj.reshape(-1)
else:
pos[:, 1] = torch.arange(N, device=device)
return pos
def _local_detail_pool(
self,
residual: torch.Tensor,
positions: Optional[torch.Tensor],
plane_ids: Optional[torch.Tensor],
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""Pool residual tokens in local coordinate windows.
Returns
-------
detail_tokens : (B, N_d_local, D)
detail_positions : (N_d_local, 4), rounded window centers
detail_counts : (N_d_local,), number of source tokens per window
"""
B, N, D = residual.shape
device = residual.device
if positions is None:
positions = self._default_positions(N, device)
positions = positions.to(device=device, dtype=torch.long)
if plane_ids is None:
plane_ids = torch.full((N,), -1, dtype=torch.long, device=device)
else:
plane_ids = plane_ids.to(device=device, dtype=torch.long)
grouped = positions.clone()
t_win = max(1, int(self.local_detail_temporal_window_size))
s_win = max(1, int(self.local_detail_window_size))
grouped[:, 0] = grouped[:, 0] // t_win
grouped[:, 1] = grouped[:, 1] // s_win
grouped[:, 2] = grouped[:, 2] // s_win
grouped[:, 3] = grouped[:, 3] // s_win
group_coords = torch.cat([plane_ids.unsqueeze(1), grouped], dim=1)
_, inverse = torch.unique(group_coords, dim=0, sorted=True, return_inverse=True)
num_groups = int(inverse.max().item()) + 1
idx = inverse.view(1, N, 1).expand(B, N, D)
pooled = residual.new_zeros(B, num_groups, D)
pooled.scatter_add_(1, idx, residual)
counts = torch.bincount(inverse, minlength=num_groups).to(device=device)
pooled = pooled / counts.view(1, num_groups, 1).clamp_min(1).to(residual.dtype)
detail_tokens = self.detail_proj(self.detail_norm(pooled))
pos_sum = torch.zeros(num_groups, 4, device=device, dtype=torch.float32)
pos_sum.scatter_add_(0, inverse.view(N, 1).expand(N, 4), positions.float())
detail_positions = (
pos_sum / counts.view(num_groups, 1).clamp_min(1).float() + 0.5
).floor().long()
return detail_tokens, detail_positions, counts
def forward(
self,
x: torch.Tensor, # (B, N, D)
positions: Optional[torch.Tensor] = None,
plane_ids: Optional[torch.Tensor] = None,
content_ratio: float = 0.25,
detail_ratio: float = 0.25,
return_metadata: bool = False,
):
"""
Returns
-------
compressed : (B, N_c + N_d_local, D)
metrics : dict with slot_diversity, residual_ratio keys
If return_metadata=True, also returns:
latent_positions : (N_c + N_d_local, 4)
latent_token_type : (N_c + N_d_local,), 0=content, 1=detail
"""
B, N, D = x.shape
N_c = max(1, int(N * content_ratio))
# Kept for pooler-key stability. Detail tokens are now determined by
# local coordinate windows rather than by global slot count.
N_d_key = max(1, int(N * detail_ratio))
content_pooler = self._get_content_pooler(N_c, N_d_key)
content_pooler = content_pooler.to(x.device)
# Stage 3a: ContentExtractor
C = content_pooler(x) # (B, N_c, D)
# Stage 3b: Residual via inverse (broadcast) attention
# weights[b, c, n] = softmax over n: sim(C[b,c], x[b,n])
weights = F.softmax(
(C @ x.transpose(-1, -2)) / (D ** 0.5), dim=-1
) # (B, N_c, N)
x_approx = weights.transpose(-1, -2) @ C # (B, N, D)
R = x - x_approx # (B, N, D)
# Stage 3c: local residual detail tokens with explicit positions
D_tokens, D_positions, detail_counts = self._local_detail_pool(
R, positions, plane_ids
)
compressed = torch.cat([C, D_tokens], dim=1) # (B, N_c + N_d, D)
# Monitoring signals
metrics = self._compute_metrics(C, R, x)
metrics['detail_token_count'] = torch.tensor(
D_tokens.shape[1], device=x.device, dtype=x.dtype)
metrics['detail_avg_window_tokens'] = detail_counts.float().mean().to(
device=x.device, dtype=x.dtype)
if not return_metadata:
return compressed, metrics
content_positions = torch.zeros(N_c, 4, dtype=torch.long, device=x.device)
latent_positions = torch.cat([content_positions, D_positions], dim=0)
latent_token_type = torch.cat([
torch.zeros(N_c, dtype=torch.long, device=x.device),
torch.ones(D_tokens.shape[1], dtype=torch.long, device=x.device),
], dim=0)
return compressed, metrics, latent_positions, latent_token_type
@staticmethod
def _compute_metrics(C: torch.Tensor, R: torch.Tensor,
x: torch.Tensor) -> Dict[str, torch.Tensor]:
with torch.no_grad():
# slot_diversity: mean pairwise cosine similarity of content slots
C_n = F.normalize(C, dim=-1) # (B, N_c, D)
sim = (C_n @ C_n.transpose(-1, -2)) # (B, N_c, N_c)
N_c = C.shape[1]
# exclude diagonal
mask = ~torch.eye(N_c, dtype=torch.bool, device=C.device)
slot_div = sim[:, mask].mean() if mask.any() else sim.mean()
# residual_ratio: ||R|| / ||x||
res_ratio = (R.norm(dim=-1) / (x.norm(dim=-1) + 1e-8)).mean()
return {'slot_diversity': slot_div, 'residual_ratio': res_ratio}
|