File size: 11,646 Bytes
49ad2ef a76a881 49ad2ef a76a881 49ad2ef a76a881 49ad2ef a76a881 49ad2ef a76a881 49ad2ef a76a881 | 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 | """Optional PyTorch Direct and Diffusion model implementations."""
from __future__ import annotations
from dataclasses import dataclass
from typing import Any
@dataclass(frozen=True)
class ModelConfig:
hidden_size: int = 256
latent_size: int = 32
diffusion_steps: int = 64
inference_steps: int = 8
abstain_threshold: float = 0.55
diffusion_conditioning: str = "token_cross_attention"
attention_heads: int = 4
def require_torch() -> tuple[Any, Any]:
try:
import torch
from torch import nn
except ImportError as exc:
raise RuntimeError("training requires optional dependencies: pip install -e '.[train]'") from exc
return torch, nn
def build_model(backend: str, *, option_count: int, config: ModelConfig | None = None) -> Any:
"""Build the decision head while keeping torch optional for data-only installs."""
_torch, nn = require_torch()
if not 2 <= option_count <= 32:
raise ValueError("option_count must be between 2 and 32")
settings = config or ModelConfig()
class DirectDecision(nn.Module):
def __init__(self) -> None:
super().__init__()
self.scorer = nn.Sequential(nn.Linear(settings.hidden_size, settings.hidden_size // 2), nn.GELU(), nn.Linear(settings.hidden_size // 2, 1))
self.answerability = nn.Linear(settings.hidden_size, 1)
def forward(self, option_embeddings: Any, pooled: Any, option_mask: Any | None = None) -> dict[str, Any]:
logits = self.scorer(option_embeddings).squeeze(-1)
if option_mask is not None:
logits = logits.masked_fill(~option_mask.bool(), float("-inf"))
return {"logits": logits, "answerability": self.answerability(pooled).squeeze(-1)}
class PooledDiffusionDecision(nn.Module):
def __init__(self) -> None:
super().__init__()
self.condition = nn.Linear(settings.hidden_size, settings.latent_size)
self.input = nn.Linear(option_count, settings.latent_size)
self.time = nn.Embedding(settings.diffusion_steps, settings.latent_size)
self.denoiser = nn.Sequential(nn.Linear(settings.latent_size, settings.hidden_size // 2), nn.SiLU(), nn.Linear(settings.hidden_size // 2, option_count))
self.answerability = nn.Linear(settings.hidden_size, 1)
def forward(self, noisy_scores: Any, pooled: Any, timestep: Any) -> dict[str, Any]:
latent = self.input(noisy_scores) + self.condition(pooled) + self.time(timestep)
return {"noise": self.denoiser(latent), "answerability": self.answerability(pooled).squeeze(-1)}
class TokenConditionedDiffusionDecision(nn.Module):
"""Denoise each option while attending to its complete token sequence."""
def __init__(self) -> None:
super().__init__()
if settings.latent_size % settings.attention_heads:
raise ValueError("latent_size must be divisible by attention_heads")
self.score_projection = nn.Linear(1, settings.latent_size)
self.token_projection = nn.Linear(settings.hidden_size, settings.latent_size)
self.time = nn.Embedding(settings.diffusion_steps, settings.latent_size)
self.cross_attention = nn.MultiheadAttention(
settings.latent_size,
settings.attention_heads,
batch_first=True,
)
self.query_norm = nn.LayerNorm(settings.latent_size)
self.denoiser = nn.Sequential(
nn.Linear(settings.latent_size, settings.hidden_size // 2),
nn.SiLU(),
nn.Linear(settings.hidden_size // 2, 1),
)
self.answerability = nn.Linear(settings.hidden_size, 1)
def forward(
self,
noisy_scores: Any,
sequence_hidden: Any,
attention_mask: Any,
timestep: Any,
option_mask: Any | None = None,
) -> dict[str, Any]:
if sequence_hidden.ndim != 4 or attention_mask.ndim != 3:
raise ValueError(
"token-conditioned diffusion expects [batch, options, sequence, hidden] states"
)
batch, options, sequence, hidden = sequence_hidden.shape
if hidden != settings.hidden_size:
raise ValueError("sequence hidden size does not match the model configuration")
flat_tokens = self.token_projection(
sequence_hidden.reshape(batch * options, sequence, hidden)
)
flat_attention = attention_mask.reshape(batch * options, sequence).bool()
time_embedding = self.time(timestep).unsqueeze(1)
queries = self.score_projection(noisy_scores.unsqueeze(-1)) + time_embedding
flat_queries = queries.reshape(batch * options, 1, settings.latent_size)
attended, _weights = self.cross_attention(
flat_queries,
flat_tokens,
flat_tokens,
key_padding_mask=~flat_attention,
need_weights=False,
)
option_logits = self.denoiser(
self.query_norm(flat_queries + attended).squeeze(1)
).reshape(batch, options)
if option_mask is not None:
option_logits = option_logits.masked_fill(~option_mask.bool(), float("-inf"))
token_weights = attention_mask.to(sequence_hidden.dtype).unsqueeze(-1)
candidate_embeddings = (sequence_hidden * token_weights).sum(dim=2) / token_weights.sum(
dim=2
).clamp_min(1)
if option_mask is None:
pooled = candidate_embeddings.mean(dim=1)
else:
option_weights = option_mask.to(sequence_hidden.dtype).unsqueeze(-1)
pooled = (candidate_embeddings * option_weights).sum(dim=1) / option_weights.sum(
dim=1
).clamp_min(1)
return {
"logits": option_logits,
"answerability": self.answerability(pooled).squeeze(-1),
}
if backend == "direct":
return DirectDecision()
if backend != "diffusion":
raise ValueError(f"unsupported backend {backend!r}")
if settings.diffusion_conditioning == "pooled":
return PooledDiffusionDecision()
if settings.diffusion_conditioning == "token_cross_attention":
return TokenConditionedDiffusionDecision()
raise ValueError(
f"unsupported diffusion_conditioning {settings.diffusion_conditioning!r}"
)
def linear_beta_schedule(steps: int) -> Any:
torch, _ = require_torch()
if steps < 2:
raise ValueError("diffusion schedule requires at least two steps")
return torch.linspace(1e-4, 0.02, steps)
def cosine_beta_schedule(steps: int, *, offset: float = 0.008) -> Any:
"""Cosine cumulative-noise schedule from Nichol and Dhariwal."""
torch, _ = require_torch()
if steps < 2:
raise ValueError("diffusion schedule requires at least two steps")
if not 0.0 < offset < 1.0:
raise ValueError("cosine offset must be between zero and one")
positions = torch.linspace(0, steps, steps + 1, dtype=torch.float64)
alpha_bars = torch.cos(((positions / steps + offset) / (1 + offset)) * torch.pi / 2) ** 2
alpha_bars = alpha_bars / alpha_bars[0]
betas = 1.0 - alpha_bars[1:] / alpha_bars[:-1]
return betas.clamp(1e-5, 0.999).to(torch.float32)
def ddim_step(sample: Any, predicted_noise: Any, alpha: Any, alpha_previous: Any) -> Any:
"""Deterministic DDIM update for a score vector."""
torch, _ = require_torch()
if sample.shape != predicted_noise.shape:
raise ValueError("sample and predicted_noise must have the same shape")
predicted_x0 = (sample - torch.sqrt(1 - alpha) * predicted_noise) / torch.sqrt(alpha)
return torch.sqrt(alpha_previous) * predicted_x0 + torch.sqrt(1 - alpha_previous) * predicted_noise
def ddim_sample(model: Any, pooled: Any, *, option_count: int, diffusion_steps: int = 32,
inference_steps: int = 4, seed: int | None = None) -> Any:
"""Sample candidate scores with deterministic DDIM updates.
The returned scores are converted to probabilities by the caller so that
candidate masking and abstention remain visible at the contract boundary.
"""
torch, _ = require_torch()
if not 2 <= inference_steps <= diffusion_steps:
raise ValueError("inference_steps must be between 2 and diffusion_steps")
betas = linear_beta_schedule(diffusion_steps).to(pooled.device)
alpha_bars = torch.cumprod(1.0 - betas, dim=0)
generator = None
if seed is not None:
generator = torch.Generator(device=pooled.device)
generator.manual_seed(seed)
sample = torch.randn((pooled.shape[0], option_count), device=pooled.device, generator=generator)
timesteps = torch.linspace(diffusion_steps - 1, 0, inference_steps, device=pooled.device).round().long()
for index, timestep in enumerate(timesteps):
current_alpha = alpha_bars[timestep]
previous_alpha = alpha_bars[timesteps[index + 1]] if index + 1 < len(timesteps) else pooled.new_tensor(1.0)
timestep_batch = torch.full((pooled.shape[0],), int(timestep), dtype=torch.long, device=pooled.device)
predicted_noise = model(sample, pooled, timestep_batch)["noise"]
sample = ddim_step(sample, predicted_noise, current_alpha, previous_alpha)
return sample
def ddim_sample_token_conditioned(
model: Any,
sequence_hidden: Any,
attention_mask: Any,
option_mask: Any,
*,
diffusion_steps: int = 64,
inference_steps: int = 8,
seed: int | None = None,
) -> Any:
"""Sample class logits from the token-conditioned x0 predictor."""
torch, _ = require_torch()
if not 2 <= inference_steps <= diffusion_steps:
raise ValueError("inference_steps must be between 2 and diffusion_steps")
batch, option_count = option_mask.shape
betas = cosine_beta_schedule(diffusion_steps).to(sequence_hidden.device)
alpha_bars = torch.cumprod(1.0 - betas, dim=0)
generator = None
if seed is not None:
generator = torch.Generator(device=sequence_hidden.device)
generator.manual_seed(seed)
sample = torch.randn(
(batch, option_count),
device=sequence_hidden.device,
generator=generator,
)
timesteps = torch.linspace(
diffusion_steps - 1,
0,
inference_steps,
device=sequence_hidden.device,
).round().long()
final_logits = sample
for index, timestep in enumerate(timesteps):
timestep_batch = torch.full(
(batch,), int(timestep), dtype=torch.long, device=sequence_hidden.device
)
output = model(
sample,
sequence_hidden,
attention_mask,
timestep_batch,
option_mask,
)
final_logits = output["logits"]
predicted_x0 = torch.softmax(final_logits, dim=-1)
current_alpha = alpha_bars[timestep]
predicted_noise = (
sample - torch.sqrt(current_alpha) * predicted_x0
) / torch.sqrt(1 - current_alpha).clamp_min(1e-6)
previous_alpha = (
alpha_bars[timesteps[index + 1]]
if index + 1 < len(timesteps)
else sample.new_tensor(1.0)
)
sample = ddim_step(sample, predicted_noise, current_alpha, previous_alpha)
return final_logits.masked_fill(~option_mask.bool(), float("-inf"))
|