File size: 12,801 Bytes
ad424e4 31604c9 ad424e4 31604c9 b233cf7 ad424e4 31604c9 b233cf7 ad424e4 b233cf7 ad424e4 b233cf7 ad424e4 82ddb20 ad424e4 b233cf7 ad424e4 b233cf7 ad424e4 557f9dd ad424e4 b233cf7 ad424e4 b233cf7 557f9dd 31604c9 557f9dd 31604c9 ad424e4 557f9dd b233cf7 ad424e4 31604c9 ad424e4 31604c9 aa64aba 31604c9 ad424e4 aa64aba b233cf7 31604c9 aa64aba 31604c9 ad424e4 b233cf7 | 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 | """PIMT classification heads — v6 ConcentrationAwarePyramidHead.
The model predicts three static pyramid tiers (top, middle, base notes) using
physics-informed concentration routing. Each tier pools token representations
weighted by OAV-based routing scores, so that top notes are dominated by
high-volatility ingredients and base notes by low-volatility ingredients.
Key design decisions (from v6 spec):
(a) Temperature-controlled routing softmax with learnable per-tier temperature.
(b) Physics-as-bias: routing = learned_attention + log_OAV / tau (init at pure physics).
(c) Scalar conditioning via 2-layer MLP (not raw concatenation).
(d) Complete padding hygiene with NaN guards.
"""
from __future__ import annotations
import json
import math
from pathlib import Path
from typing import Any
import torch
import torch.nn as nn
import torch.nn.functional as F
def _load_scalar_stats(path: str = "artifacts/scalar_stats_v1.json") -> dict[str, dict[str, float]]:
"""Load training-set scalar statistics for standardization."""
p = Path(path)
if p.exists():
return json.loads(p.read_text())
# Fallback defaults
return {
"log10_oav_sum": {"mean": 35.0, "std": 21.5},
"x_liquid_sum": {"mean": 1.0, "std": 0.08},
}
class ConcentrationAwarePyramidHead(nn.Module):
"""Three-tier pyramid head with physics-informed concentration routing.
Input:
latent: (B, S, H) — per-token hidden states from the transformer (time already pooled).
physics: (B, S, 2) — per-token physics states [x_liquid, log10(OAV)].
src_key_padding_mask: (B, S) — True for padding positions.
Output: (B, 3, 138) sigmoid probabilities for [top, mid, base].
"""
def __init__(
self,
hidden_dim: int,
output_dim: int = 138,
scalar_stats: dict[str, dict[str, float]] | None = None,
scalar_embed_dim: int = 16,
) -> None:
super().__init__()
self.hidden_dim = hidden_dim
self.output_dim = output_dim
# Load scalar standardization constants
if scalar_stats is None:
scalar_stats = _load_scalar_stats()
oav_stats = scalar_stats.get("log10_oav_sum", {"mean": 35.0, "std": 21.5})
xliq_stats = scalar_stats.get("x_liquid_sum", {"mean": 1.0, "std": 0.08})
self.register_buffer("oav_mean", torch.tensor(oav_stats["mean"], dtype=torch.float32))
self.register_buffer("oav_std", torch.tensor(max(oav_stats["std"], 1e-6), dtype=torch.float32))
self.register_buffer("xliq_mean", torch.tensor(xliq_stats["mean"], dtype=torch.float32))
self.register_buffer("xliq_std", torch.tensor(max(xliq_stats["std"], 1e-6), dtype=torch.float32))
# (a) Learnable per-tier temperature for routing softmax.
# Init tau=1.0, clamp to [0.1, 10].
self.log_tau = nn.Parameter(torch.zeros(3)) # log(1.0) = 0
# (b) Physics-as-bias: learned attention projection (init at zero = pure physics).
self.attn_proj = nn.Linear(hidden_dim, 1, bias=False)
nn.init.zeros_(self.attn_proj.weight) # Start at pure-physics routing
# (c) Scalar conditioning MLP: [standardized log10(ΣOAV), standardized Σx_liquid] → 16-D.
self.scalar_mlp = nn.Sequential(
nn.Linear(2, scalar_embed_dim),
nn.GELU(),
nn.Linear(scalar_embed_dim, scalar_embed_dim),
)
# Tier-specific linear heads that take the routed pooled vector + scalar embedding.
pooled_dim = hidden_dim + scalar_embed_dim
self.top_head = nn.Linear(pooled_dim, output_dim)
self.mid_head = nn.Linear(pooled_dim, output_dim)
self.base_head = nn.Linear(pooled_dim, output_dim)
# Tier-specific scalar stats for routing (which OAV snapshot to use)
# We use three time-window OAV summaries. For now, they share the overall stats.
# The routing is per-token: each token gets a routing weight per tier.
def _compute_routing_weights(
self,
latent: torch.Tensor,
physics: torch.Tensor,
src_key_padding_mask: torch.Tensor,
) -> torch.Tensor:
"""Compute per-tier routing weights for each token.
Args:
latent: (B, S, H)
physics: (B, S, 2) — [x_liquid, log10(OAV)]
src_key_padding_mask: (B, S) — True for padding
Returns:
routing_weights: (B, 3, S) — per-tier softmax weights over tokens.
"""
B, S, H = latent.shape
# Extract log10(OAV) per token (physics channel 1)
log_oav = physics[..., 1] # (B, S)
# (b) Learned attention score (init at zero → pure physics at start)
attn_score = self.attn_proj(latent).squeeze(-1) # (B, S)
# Temperature: tau = exp(clamp(log_tau, -2.3, 2.3)) → clamp tau to [0.1, 10]
tau = torch.exp(torch.clamp(self.log_tau, -2.303, 2.303)) # (3,)
routing_weights = []
for tier_idx in range(3):
# Routing logits = w * attn_score + log_oav / tau
# w is folded into attn_score (single learned weight per tier)
# For simplicity, attn_score is shared but tau differs per tier
logits = attn_score + log_oav / tau[tier_idx] # (B, S)
# (d) Mask padding with finfo.min (not hardcoded -1e9)
neg_mask_val = torch.finfo(logits.dtype).min
logits = logits.masked_fill(src_key_padding_mask, neg_mask_val)
# NaN guard: if ALL positions are masked (shouldn't happen), use uniform
all_masked = src_key_padding_mask.all(dim=1, keepdim=True) # (B, 1)
if all_masked.any():
# Replace fully-masked rows with uniform distribution
safe_logits = torch.zeros_like(logits)
safe_logits = safe_logits.masked_fill(src_key_padding_mask, neg_mask_val)
logits = torch.where(all_masked.expand_as(logits), safe_logits, logits)
weights = F.softmax(logits, dim=-1) # (B, S)
routing_weights.append(weights)
return torch.stack(routing_weights, dim=1) # (B, 3, S)
def _compute_intensity_scalars(
self,
physics: torch.Tensor,
src_key_padding_mask: torch.Tensor,
) -> torch.Tensor:
"""Compute standardized intensity scalars: [log10(ΣOAV), Σx_liquid].
These are computed per-tier (using different OAV snapshots), but for now
we use the overall per-token physics.
Args:
physics: (B, S, 2) — [x_liquid, log10(OAV)]
src_key_padding_mask: (B, S)
Returns:
scalars: (B, 2) — standardized [log10(ΣOAV), Σx_liquid]
"""
B, S, _ = physics.shape
# Zero out padding in the sums
mask = (~src_key_padding_mask).float().unsqueeze(-1) # (B, S, 1)
masked_physics = physics * mask # (B, S, 2)
# Sum over tokens (axis 1)
log_oav_sum = masked_physics[..., 1].sum(dim=1) # (B,)
x_liq_sum = masked_physics[..., 0].sum(dim=1) # (B,)
# Standardize using training-set statistics
std_oav = (log_oav_sum - self.oav_mean) / self.oav_std # (B,)
std_xliq = (x_liq_sum - self.xliq_mean) / self.xliq_std # (B,)
return torch.stack([std_oav, std_xliq], dim=-1) # (B, 2)
def forward(
self,
latent: torch.Tensor,
physics: torch.Tensor,
src_key_padding_mask: torch.Tensor,
) -> dict[str, torch.Tensor]:
"""Forward pass.
Args:
latent: (B, S, H) — per-token hidden states.
physics: (B, S, 2) — [x_liquid, log10(OAV)] per token.
src_key_padding_mask: (B, S) — True for padding.
Returns:
dict with:
"pyramid": (B, 3, 138) sigmoid probabilities
"routing_weights": (B, 3, S) routing weights (for diagnostics)
"scalar_embedding": (B, 16) scalar conditioning embedding
"""
B, S, H = latent.shape
# (a, b) Compute routing weights
routing = self._compute_routing_weights(latent, physics, src_key_padding_mask) # (B, 3, S)
# (c) Compute scalar conditioning
scalars = self._compute_intensity_scalars(physics, src_key_padding_mask) # (B, 2)
scalar_emb = self.scalar_mlp(scalars) # (B, 16)
# Routed pooling: weighted sum of token representations per tier
# routing: (B, 3, S), latent: (B, S, H)
pooled = torch.bmm(routing, latent) # (B, 3, H)
# Concatenate scalar embedding to each tier
scalar_expanded = scalar_emb.unsqueeze(1).expand(-1, 3, -1) # (B, 3, 16)
pooled_with_scalar = torch.cat([pooled, scalar_expanded], dim=-1) # (B, 3, H+16)
# Tier-specific linear heads + sigmoid
top = torch.sigmoid(self.top_head(pooled_with_scalar[:, 0])) # (B, 138)
mid = torch.sigmoid(self.mid_head(pooled_with_scalar[:, 1]))
base = torch.sigmoid(self.base_head(pooled_with_scalar[:, 2]))
pyramid = torch.stack([top, mid, base], dim=1) # (B, 3, 138)
return {
"pyramid": pyramid,
"routing_weights": routing,
"scalar_embedding": scalar_emb,
}
class PIMTHeads(nn.Module):
"""Container for the concentration-aware pyramid head and subjective heads.
Accepts the full 4D latent from the transformer (B, T, S, H) and pools
over time before feeding to the pyramid head.
"""
def __init__(
self,
hidden_dim: int,
objective_dim: int = 138,
seasonality_classes: int = 4,
wearability_classes: int = 2,
scalar_stats: dict[str, dict[str, float]] | None = None,
) -> None:
super().__init__()
self.objective_head = ConcentrationAwarePyramidHead(
hidden_dim, objective_dim, scalar_stats=scalar_stats
)
# Subjective heads operate on a globally-pooled representation.
self.seasonality_head = nn.Linear(hidden_dim, seasonality_classes)
self.gender_head = nn.Linear(hidden_dim, 1)
self.wearability_head = nn.Linear(hidden_dim, wearability_classes)
self.substantivity_head = nn.Linear(hidden_dim, 1)
def forward(
self,
latent: torch.Tensor,
physics: torch.Tensor | None = None,
src_key_padding_mask: torch.Tensor | None = None,
) -> dict[str, Any]:
"""Forward pass.
Args:
latent: (B, T, S, H) from the transformer encoder.
physics: (B, T, S, 2) physics states. If None, uses zero physics.
src_key_padding_mask: (B, S) padding mask. If None, no padding.
Returns:
dict with objective, subjective, alignment, and diagnostics.
"""
# Pool over time dimension to get (B, S, H)
latent_pooled_time = latent.mean(dim=1) # (B, S, H)
B, S, H = latent_pooled_time.shape
# Handle physics input
if physics is not None:
# Pool physics over time too: (B, T, S, 2) → (B, S, 2)
physics_pooled = physics.mean(dim=1) # (B, S, 2)
else:
physics_pooled = latent_pooled_time.new_zeros(B, S, 2)
# Handle padding mask
if src_key_padding_mask is None:
src_key_padding_mask = torch.zeros(B, S, dtype=torch.bool, device=latent.device)
# Concentration-aware pyramid prediction
pyramid_result = self.objective_head(latent_pooled_time, physics_pooled, src_key_padding_mask)
pyramid = pyramid_result["pyramid"] # (B, 3, 138)
# Alignment head: mean of tier logits for cosine alignment loss
pooled_global = latent_pooled_time.mean(dim=1) # (B, H)
alignment_raw = pyramid_result["pyramid"].mean(dim=1) # (B, 138) — already sigmoided
alignment = alignment_raw # Use the mean of sigmoid outputs as alignment target
# Subjective heads
seasonality = self.seasonality_head(pooled_global)
gender = self.gender_head(pooled_global)
wearability = self.wearability_head(pooled_global)
substantivity = self.substantivity_head(pooled_global)
return {
"objective": pyramid,
"subjective": {
"seasonality": seasonality,
"gender_profile": gender,
"wearability": wearability,
"substantivity": substantivity,
},
"alignment": alignment,
"diagnostics": {
"routing_weights": pyramid_result["routing_weights"],
"scalar_embedding": pyramid_result["scalar_embedding"],
"tau": torch.exp(torch.clamp(self.objective_head.log_tau, -2.303, 2.303)),
},
}
|