File size: 12,480 Bytes
9463e5c | 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 | """
GLADIUS — Gaussian Specialist Head
The specialist module that plugs into WYRM's NexusRouter.
Generates 3D Gaussian Splat parameters from backbone hidden states.
Two-stage hierarchical generation:
Stage 1 (Anchors): Direct regression of coarse Gaussians from pooled hidden state
Stage 2 (Details): VQ-coded fine Gaussians via cross-attention to backbone features
Depth profile integration: learned per-layer gates determine which backbone
layers contribute to structure (anchors) vs detail.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from .config import GaussianConfig
from .vqvae import GaussianVQVAE, GaussianVQDecoder
class AnchorHead(nn.Module):
"""
Generates K coarse anchor Gaussians from the backbone's pooled representation.
Each anchor: position(3) + scale(3) + rotation(4) + opacity(1) + sh_dc(3) = 14 floats.
Uses direct regression — anchors need precise continuous placement.
"""
def __init__(self, backbone_dim: int, config: GaussianConfig):
super().__init__()
self.config = config
self.n_anchors = config.num_anchors
self.out_dim = config.full_dim # 14
self.net = nn.Sequential(
nn.Linear(backbone_dim, config.anchor_hidden),
nn.LayerNorm(config.anchor_hidden),
nn.GELU(),
nn.Linear(config.anchor_hidden, config.anchor_hidden),
nn.LayerNorm(config.anchor_hidden),
nn.GELU(),
nn.Linear(config.anchor_hidden, self.n_anchors * self.out_dim),
)
self._init_weights()
def _init_weights(self):
# Initialize last layer small — anchors start near origin
nn.init.normal_(self.net[-1].weight, std=0.01)
nn.init.zeros_(self.net[-1].bias)
def forward(self, pooled: torch.Tensor) -> torch.Tensor:
"""
Args:
pooled: (B, backbone_dim) — pooled backbone output
Returns:
anchors: (B, K, 14) — K anchor Gaussians with full params
"""
B = pooled.shape[0]
raw = self.net(pooled) # (B, K*14)
raw = raw.view(B, self.n_anchors, self.out_dim)
# Activate each component appropriately
anchors = self._activate(raw)
return anchors
def _activate(self, raw: torch.Tensor) -> torch.Tensor:
"""Apply per-component activations to raw output."""
pos = torch.tanh(raw[..., :3]) * self.config.scene_scale # [-scale, +scale]
scale = raw[..., 3:6].clamp(self.config.min_gaussian_scale,
self.config.max_gaussian_scale) # Log-scale clamped
rot = F.normalize(raw[..., 6:10], dim=-1) # Unit quaternion
opacity = torch.sigmoid(raw[..., 10:11]) # [0, 1]
color = torch.sigmoid(raw[..., 11:14]) # [0, 1] RGB
return torch.cat([pos, scale, rot, opacity, color], dim=-1)
class DetailHead(nn.Module):
"""
Generates M fine detail Gaussians per anchor using VQ tokens.
Cross-attends from anchor queries to backbone hidden states,
then predicts VQ codebook indices + continuous position offsets.
"""
def __init__(self, backbone_dim: int, config: GaussianConfig):
super().__init__()
self.config = config
self.n_details = config.details_per_anchor
self.backbone_dim = backbone_dim
# Anchor → query expansion: each anchor generates M query vectors
self.query_expand = nn.Sequential(
nn.Linear(config.full_dim, config.detail_hidden),
nn.GELU(),
nn.Linear(config.detail_hidden, self.n_details * backbone_dim),
)
# Cross-attention: detail queries attend to backbone features
self.cross_attn = nn.MultiheadAttention(
embed_dim=backbone_dim,
num_heads=config.cross_attn_heads,
batch_first=True,
dropout=0.0,
)
self.cross_norm = nn.LayerNorm(backbone_dim)
# VQ index prediction
self.vq_proj = nn.Sequential(
nn.Linear(backbone_dim, backbone_dim),
nn.GELU(),
nn.Linear(backbone_dim, config.codebook_size),
)
# Continuous position offset (relative to anchor)
self.offset_proj = nn.Sequential(
nn.Linear(backbone_dim, config.detail_hidden),
nn.GELU(),
nn.Linear(config.detail_hidden, 3),
)
def forward(self, anchors: torch.Tensor, backbone_features: torch.Tensor) -> dict:
"""
Args:
anchors: (B, K, 14) — anchor Gaussians
backbone_features: (B, S, backbone_dim) — backbone hidden states
Returns:
dict with vq_logits, vq_indices, pos_offsets
"""
B, K, _ = anchors.shape
# Expand each anchor into M query vectors
queries = self.query_expand(anchors) # (B, K, M * D)
queries = queries.view(B, K * self.n_details, self.backbone_dim) # (B, K*M, D)
# Cross-attend to backbone
detail_features, _ = self.cross_attn(
queries, backbone_features, backbone_features
) # (B, K*M, D)
detail_features = self.cross_norm(detail_features + queries) # Residual
# Predict VQ indices
vq_logits = self.vq_proj(detail_features) # (B, K*M, codebook_size)
vq_indices = vq_logits.argmax(dim=-1) # (B, K*M)
# Predict position offsets
pos_offsets = self.offset_proj(detail_features) # (B, K*M, 3)
# Scale offsets — details should be close to their anchor
pos_offsets = torch.tanh(pos_offsets) * 0.5 # [-0.5, +0.5] around anchor
return {
'vq_logits': vq_logits,
'vq_indices': vq_indices,
'pos_offsets': pos_offsets,
'detail_features': detail_features,
}
class GaussianSpecialist(nn.Module):
"""
The complete Gaussian specialist head for WYRM.
Plugs into NexusRouter as specialist index N.
Generates 3D Gaussian splat scenes from backbone hidden states.
Two-stage generation:
1. Anchor Head → K coarse Gaussians (direct regression)
2. Detail Head → K*M fine Gaussians (VQ-coded)
Depth profile integration: learned per-layer gates select which
backbone layers contribute to structure vs detail.
"""
def __init__(self, backbone_dim: int, num_backbone_layers: int,
config: GaussianConfig, vqvae: GaussianVQVAE = None):
super().__init__()
self.config = config
self.backbone_dim = backbone_dim
self.num_layers = num_backbone_layers
# ── Sub-modules ──
self.anchor_head = AnchorHead(backbone_dim, config)
self.detail_head = DetailHead(backbone_dim, config)
# ── VQ-VAE decoder (frozen — pre-trained in Phase 1) ──
if vqvae is not None:
self.vq_decoder = vqvae.decoder
# Freeze VQ decoder
for p in self.vq_decoder.parameters():
p.requires_grad = False
# Also keep the quantizer's codebook for decoding
self.register_buffer('vq_codebook', vqvae.quantizer.embed.clone())
else:
self.vq_decoder = None
self.vq_codebook = None
# ── Depth Profile Gates ──
# Learned: which backbone layers matter for anchors vs details
self.anchor_layer_gate = nn.Parameter(torch.zeros(num_backbone_layers))
self.detail_layer_gate = nn.Parameter(torch.zeros(num_backbone_layers))
# ── Projection for pooling ──
self.pool_proj = nn.Linear(backbone_dim, backbone_dim)
def forward(self, layer_outputs: list[torch.Tensor]) -> dict:
"""
Args:
layer_outputs: List of (B, S, D) tensors from each backbone layer.
Returns:
dict with:
anchors: (B, K, 14) — anchor Gaussians
details: (B, K*M, 14) — detail Gaussians (if VQ decoder available)
all_gaussians: (B, K + K*M, 14) — concatenated scene
vq_logits: (B, K*M, codebook_size) — for training loss
pos_offsets: (B, K*M, 3) — detail position offsets
"""
B = layer_outputs[0].shape[0]
# ── Depth-profiled aggregation for anchors ──
anchor_weights = torch.softmax(self.anchor_layer_gate, dim=0)
anchor_hidden = sum(
w * h for w, h in zip(anchor_weights, layer_outputs)
) # (B, S, D)
# Pool → single vector per batch
pooled = self.pool_proj(anchor_hidden.mean(dim=1)) # (B, D)
# ── Stage 1: Generate anchors ──
anchors = self.anchor_head(pooled) # (B, K, 14)
# ── Depth-profiled aggregation for details ──
detail_weights = torch.softmax(self.detail_layer_gate, dim=0)
detail_hidden = sum(
w * h for w, h in zip(detail_weights, layer_outputs)
) # (B, S, D)
# ── Stage 2: Generate details ──
detail_out = self.detail_head(anchors, detail_hidden)
result = {
'anchors': anchors,
'vq_logits': detail_out['vq_logits'],
'vq_indices': detail_out['vq_indices'],
'pos_offsets': detail_out['pos_offsets'],
}
# ── Decode VQ indices to full Gaussian params (if decoder available) ──
if self.vq_decoder is not None and self.vq_codebook is not None:
details = self._decode_details(
anchors, detail_out['vq_indices'], detail_out['pos_offsets']
)
result['details'] = details
result['all_gaussians'] = torch.cat([anchors, details], dim=1)
return result
def _decode_details(self, anchors: torch.Tensor, vq_indices: torch.Tensor,
pos_offsets: torch.Tensor) -> torch.Tensor:
"""
Decode VQ indices + anchor positions → full detail Gaussians.
Args:
anchors: (B, K, 14)
vq_indices: (B, K*M)
pos_offsets: (B, K*M, 3)
Returns:
details: (B, K*M, 14) — fully parameterized detail Gaussians
"""
B, KM = vq_indices.shape
K = self.config.num_anchors
M = self.config.details_per_anchor
# Decode VQ → (scale, rot, opacity, color)
z_q = self.vq_codebook[vq_indices.view(-1)] # (B*K*M, codebook_dim)
params = self.vq_decoder(z_q) # (B*K*M, param_dim=11)
params = params.view(B, KM, self.config.param_dim)
# Extract components
scale = params[..., :3].clamp(self.config.min_gaussian_scale,
self.config.max_gaussian_scale)
rot = F.normalize(params[..., 3:7], dim=-1)
opacity = torch.sigmoid(params[..., 7:8])
color = torch.sigmoid(params[..., 8:11])
# Compute world-space positions: anchor_pos + offset
anchor_pos = anchors[:, :, :3] # (B, K, 3)
# Repeat each anchor M times
anchor_pos_expanded = anchor_pos.unsqueeze(2).expand(B, K, M, 3).reshape(B, KM, 3)
world_pos = anchor_pos_expanded + pos_offsets
return torch.cat([world_pos, scale, rot, opacity, color], dim=-1)
def get_depth_profile(self) -> dict:
"""Return the learned depth profile weights for analysis."""
with torch.no_grad():
anchor_w = torch.softmax(self.anchor_layer_gate, dim=0)
detail_w = torch.softmax(self.detail_layer_gate, dim=0)
return {
'anchor_weights': anchor_w.cpu().tolist(),
'detail_weights': detail_w.cpu().tolist(),
'anchor_peak_layer': anchor_w.argmax().item(),
'detail_peak_layer': detail_w.argmax().item(),
}
def count_params(self) -> dict:
"""Count parameters by component."""
def count(module):
return sum(p.numel() for p in module.parameters() if p.requires_grad)
return {
'anchor_head': count(self.anchor_head),
'detail_head': count(self.detail_head),
'pool_proj': count(self.pool_proj),
'layer_gates': self.anchor_layer_gate.numel() + self.detail_layer_gate.numel(),
'total': sum(p.numel() for p in self.parameters() if p.requires_grad),
}
|