File size: 6,600 Bytes
358d3bc | 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 | """
ProbVLM-Style Probabilistic Adapter for Uncertainty Estimation.
Converts point embeddings into distributions (Generalized Gaussian)
following the BayesCap approach from ProbVLM.
Each adapter takes a frozen embedding and predicts:
mu: Shift from the input embedding (residual)
alpha: Scale parameter (controls spread)
beta: Shape parameter (controls tail behavior)
These define a Generalized Gaussian distribution:
p(x) ∝ exp(-(|x - mu| / alpha)^beta)
MC sampling from this distribution produces N embedding samples,
which propagate uncertainty through the Gramian volume computation.
Architecture: BayesCap_MLP
input → Linear(d, hidden) → ReLU → Dropout
→ Linear(hidden, hidden) → ReLU → Dropout
→ Three heads: mu_head, alpha_head, beta_head
"""
from __future__ import annotations
import logging
from pathlib import Path
from typing import Dict, Optional, Tuple
import numpy as np
logger = logging.getLogger(__name__)
try:
import torch
import torch.nn as nn
import torch.nn.functional as F
TORCH_AVAILABLE = True
except ImportError:
TORCH_AVAILABLE = False
def _check_torch():
if not TORCH_AVAILABLE:
raise ImportError("PyTorch required for ProbabilisticAdapter")
class ProbabilisticAdapter(nn.Module):
"""
BayesCap-style adapter that maps point embeddings to distributions.
Takes a frozen embedding (from CLIP or CLAP) and predicts
Generalized Gaussian parameters: (mu, alpha, beta).
The adapter is lightweight (~0.5M params) and trains in minutes
on small datasets.
"""
def __init__(
self,
input_dim: int = 512,
hidden_dim: int = 256,
num_layers: int = 3,
dropout: float = 0.1,
):
_check_torch()
super().__init__()
self.input_dim = input_dim
# Shared backbone
layers = []
in_d = input_dim
for _ in range(num_layers - 1):
layers.extend([
nn.Linear(in_d, hidden_dim),
nn.ReLU(),
nn.Dropout(dropout),
])
in_d = hidden_dim
self.backbone = nn.Sequential(*layers)
# Three output heads
self.mu_head = nn.Linear(hidden_dim, input_dim)
self.alpha_head = nn.Linear(hidden_dim, input_dim)
self.beta_head = nn.Linear(hidden_dim, input_dim)
self.config = {
"input_dim": input_dim,
"hidden_dim": hidden_dim,
"num_layers": num_layers,
"dropout": dropout,
}
def forward(
self, embedding: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Predict distribution parameters from a point embedding.
Args:
embedding: Input embedding [batch, input_dim].
Returns:
mu: Location parameter [batch, input_dim] (embedding + residual)
alpha: Scale parameter [batch, input_dim] (> 0, via softplus)
beta: Shape parameter [batch, input_dim] (> 0, via softplus)
"""
h = self.backbone(embedding)
# mu: residual + input (anchored to original embedding)
mu = embedding + self.mu_head(h)
# alpha, beta: positive via softplus
alpha = F.softplus(self.alpha_head(h)) + 1e-6
beta = F.softplus(self.beta_head(h)) + 1e-6
return mu, alpha, beta
def sample(
self,
embedding: np.ndarray,
n_samples: int = 100,
) -> np.ndarray:
"""
Draw Monte Carlo samples from the predicted distribution.
Uses the reparameterization trick for Generalized Gaussian:
x = mu + alpha * sign(u) * |u|^(1/beta)
where u ~ Uniform(-1, 1)
Args:
embedding: Input embedding, shape (dim,) or (1, dim).
n_samples: Number of MC samples.
Returns:
Samples array, shape (n_samples, dim).
"""
_check_torch()
self.eval()
emb = embedding.squeeze()
if emb.ndim == 1:
emb = emb[np.newaxis, :]
with torch.no_grad():
x = torch.tensor(emb, dtype=torch.float32)
mu, alpha, beta = self.forward(x)
# Expand for sampling: [1, dim] -> [n_samples, dim]
mu = mu.expand(n_samples, -1)
alpha = alpha.expand(n_samples, -1)
beta = beta.expand(n_samples, -1)
# Reparameterized sampling from Generalized Gaussian
u = torch.rand_like(mu) * 2 - 1 # Uniform(-1, 1)
sign = torch.sign(u)
samples = mu + alpha * sign * (torch.abs(u) + 1e-8).pow(1.0 / beta)
# L2 normalize samples (stay on unit sphere)
samples = F.normalize(samples, p=2, dim=-1)
return samples.cpu().numpy()
def uncertainty(self, embedding: np.ndarray) -> float:
"""
Compute scalar aleatoric uncertainty for an embedding.
Returns the mean predicted alpha (scale parameter) across dimensions.
High alpha → high uncertainty → wide distribution.
Args:
embedding: Input embedding, shape (dim,) or (1, dim).
Returns:
Scalar uncertainty value (mean alpha).
"""
_check_torch()
self.eval()
emb = embedding.squeeze()
if emb.ndim == 1:
emb = emb[np.newaxis, :]
with torch.no_grad():
x = torch.tensor(emb, dtype=torch.float32)
_, alpha, _ = self.forward(x)
return float(alpha.mean().item())
def save(self, path: str) -> None:
"""Save adapter weights + config."""
_check_torch()
import json
p = Path(path)
p.parent.mkdir(parents=True, exist_ok=True)
torch.save(self.state_dict(), p)
config_path = p.with_suffix(".json")
with config_path.open("w") as f:
json.dump(self.config, f, indent=2)
logger.info("Saved ProbabilisticAdapter to %s", path)
@classmethod
def load(cls, path: str) -> "ProbabilisticAdapter":
"""Load adapter from saved weights."""
_check_torch()
import json
p = Path(path)
config_path = p.with_suffix(".json")
with config_path.open("r") as f:
config = json.load(f)
model = cls(**config)
state_dict = torch.load(p, map_location="cpu", weights_only=True)
model.load_state_dict(state_dict)
model.eval()
logger.info("Loaded ProbabilisticAdapter from %s", path)
return model
|