File size: 10,050 Bytes
d9d4242 100bc30 d9d4242 | 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 | """DDIM sampling with classifier-free guidance over CUE embeddings."""
from __future__ import annotations
from dataclasses import dataclass
from pathlib import Path
from typing import Any
import numpy as np
import torch
from .sampler_diffusion import CosineNoiseSchedule, DiffusionSamplerModel
SAMPLER_NAME = "sampler.pt"
def project_layernorm_manifold(
z: torch.Tensor,
weight: torch.Tensor | None,
bias: torch.Tensor | None,
eps: float = 1e-5,
) -> torch.Tensor:
"""Exact re-application of encoder bottleneck LayerNorm affine."""
if weight is None or bias is None:
# Unit-sphere-ish recenter/rescale fallback.
mean = z.mean(dim=-1, keepdim=True)
var = z.var(dim=-1, keepdim=True, unbiased=False)
return (z - mean) / torch.sqrt(var + eps)
# Invert y = gamma * x_hat + beta => x_hat = (y - beta) / gamma
x_hat = (z - bias) / weight.clamp_min(eps)
# Re-center/re-scale then reapply affine (projects onto LN manifold).
mean = x_hat.mean(dim=-1, keepdim=True)
var = x_hat.var(dim=-1, keepdim=True, unbiased=False)
x_hat = (x_hat - mean) / torch.sqrt(var + eps)
return x_hat * weight + bias
@torch.no_grad()
def ddim_sample(
model: DiffusionSamplerModel,
schedule: CosineNoiseSchedule,
*,
n: int,
steps: int = 50,
guidance_w: float = 1.5,
set_emb: torch.Tensor | None = None,
mask: torch.Tensor | None = None,
generator: torch.Generator | None = None,
device: torch.device | str = "cpu",
) -> torch.Tensor:
"""
Sample standardized embeddings via DDIM + CFG.
set_emb: [B, K, D] standardized condition set (or None for unconditional).
guidance_w: 0 / no set -> unconditional; typical 1–3 for conditional.
"""
model.eval()
device = torch.device(device)
dim = model.dim
if set_emb is not None:
batch = set_emb.shape[0]
if n != batch:
# allow n to override only when broadcasting a single set
if batch == 1 and n > 1:
set_emb = set_emb.expand(n, -1, -1)
if mask is not None:
mask = mask.expand(n, -1)
else:
n = batch
zt = torch.randn(n, dim, device=device, generator=generator)
timesteps = torch.linspace(schedule.timesteps - 1, 0, steps, device=device).long()
cond = None
null = model.conditioner.null.unsqueeze(0).expand(n, -1)
if set_emb is not None and guidance_w != 0:
cond = model.conditioner(set_emb, mask)
for i, t_scalar in enumerate(timesteps):
t = torch.full((n,), int(t_scalar.item()), device=device, dtype=torch.long)
eps_uncond = model.denoiser(zt, t, null)
if cond is None or guidance_w == 0:
eps = eps_uncond
else:
eps_cond = model.denoiser(zt, t, cond)
eps = eps_uncond + guidance_w * (eps_cond - eps_uncond)
x0 = schedule.predict_x0_from_eps(zt, t, eps)
if i + 1 < len(timesteps):
t_next = torch.full(
(n,), int(timesteps[i + 1].item()), device=device, dtype=torch.long
)
alpha_next = schedule._gather(schedule.alphas_cumprod, t_next, like=zt)
# DDIM (eta=0): z_{t-1} = sqrt(a_next)*x0 + sqrt(1-a_next)*eps
zt = torch.sqrt(alpha_next) * x0 + torch.sqrt(1.0 - alpha_next) * eps
else:
zt = x0
return zt
@dataclass
class CueSampler:
"""Loaded diffusion prior over CUE bottlenecks, with its standardization stats."""
model: DiffusionSamplerModel
schedule: CosineNoiseSchedule
mu: torch.Tensor
sigma: torch.Tensor
layernorm_weight: torch.Tensor | None
layernorm_bias: torch.Tensor | None
config: dict[str, Any]
@property
def dim(self) -> int:
return int(self.model.dim)
def to(self, device: str | torch.device) -> CueSampler:
self.model.to(device)
self.schedule.to(device)
self.mu = self.mu.to(device)
self.sigma = self.sigma.to(device)
if self.layernorm_weight is not None:
self.layernorm_weight = self.layernorm_weight.to(device)
if self.layernorm_bias is not None:
self.layernorm_bias = self.layernorm_bias.to(device)
return self
@torch.no_grad()
def sample(
self,
*,
n: int = 8,
condition: np.ndarray | torch.Tensor | list | None = None,
steps: int | None = None,
guidance_w: float | None = None,
project_manifold: bool | None = None,
seed: int = 0,
) -> torch.Tensor:
"""Return unstandardized CUE embeddings [n, D]."""
device = self.mu.device
steps = int(steps if steps is not None else self.config.get("ddim_steps", 50))
if guidance_w is None:
guidance_w = float(self.config.get("guidance_w", 1.5))
if project_manifold is None:
project_manifold = bool(self.config.get("project_manifold", False))
set_emb = None
mask = None
if condition is not None:
cond = as_condition_tensor(condition, device=device)
set_emb = (cond - self.mu) / self.sigma
mask = torch.ones(
set_emb.shape[0], set_emb.shape[1], dtype=torch.bool, device=device
)
if not torch.isfinite(set_emb).all():
raise ValueError("condition embeddings contain non-finite values")
if set_emb.shape[-1] != self.dim:
raise ValueError(f"expected dim={self.dim}, got {set_emb.shape[-1]}")
generator = torch.Generator(device=device)
generator.manual_seed(int(seed))
z_std = ddim_sample(
self.model,
self.schedule,
n=n,
steps=steps,
guidance_w=float(guidance_w) if set_emb is not None else 0.0,
set_emb=set_emb,
mask=mask,
generator=generator,
device=device,
)
z = z_std * self.sigma + self.mu
if project_manifold:
z = project_layernorm_manifold(
z, self.layernorm_weight, self.layernorm_bias
)
return z
def load_sampler(
checkpoint: str | Path,
*,
device: str = "cpu",
use_ema: bool = True,
) -> CueSampler:
"""Load ``sampler.pt`` from a CUE sampler directory (or a direct file path)."""
path = Path(checkpoint)
if path.is_dir():
path = path / SAMPLER_NAME
if not path.is_file():
raise FileNotFoundError(f"sampler weights not found: {path}")
payload = torch.load(path, map_location=device, weights_only=False)
cfg = payload.get("config") or {}
dim = int(cfg.get("dim", payload.get("dim", 1024)))
model = DiffusionSamplerModel(
dim=dim,
width=int(cfg.get("width", 1536)),
depth=int(cfg.get("depth", 6)),
time_dim=int(cfg.get("time_dim", 256)),
conditioner_heads=int(cfg.get("conditioner_heads", 4)),
).to(device)
if use_ema and "ema" in payload:
model.load_state_dict(payload["ema"]["shadow"])
else:
model.load_state_dict(payload["model"])
model.eval()
schedule = CosineNoiseSchedule.from_state_dict(payload["schedule"]).to(device)
ln_w = payload.get("layernorm_weight")
ln_b = payload.get("layernorm_bias")
return CueSampler(
model=model,
schedule=schedule,
mu=torch.tensor(payload["mu"], device=device, dtype=torch.float32),
sigma=torch.tensor(payload["sigma"], device=device, dtype=torch.float32),
layernorm_weight=(
torch.tensor(ln_w, device=device, dtype=torch.float32)
if ln_w is not None
else None
),
layernorm_bias=(
torch.tensor(ln_b, device=device, dtype=torch.float32)
if ln_b is not None
else None
),
config=dict(cfg),
)
"""Keys ``load_sampler`` reads, plus provenance scalars. Everything else is training state."""
INFERENCE_KEYS = (
"schedule",
"mu",
"sigma",
"layernorm_weight",
"layernorm_bias",
"dim",
"config",
"checkpoint_hash",
"step",
)
def slim_sampler_payload(payload: dict[str, Any]) -> dict[str, Any]:
"""Drop optimizer/scheduler/scaler state and the duplicate weight copy.
Training checkpoints hold ``model``, an EMA shadow, and optimizer moments. Inference
uses the EMA weights, so those are promoted into ``model`` and the rest is dropped:
one copy that loads identically whether or not ``use_ema`` is set.
"""
ema = payload.get("ema")
if isinstance(ema, dict) and "shadow" in ema:
weights = ema["shadow"]
elif "model" in payload:
weights = payload["model"]
else:
raise KeyError("sampler payload has neither ema['shadow'] nor model weights")
slim: dict[str, Any] = {"model": weights}
slim.update({key: payload[key] for key in INFERENCE_KEYS if key in payload})
return slim
def slim_sampler_file(source: str | Path, destination: str | Path) -> Path:
"""Rewrite a ``sampler.pt`` with only the tensors inference needs."""
source, destination = Path(source), Path(destination)
if source.is_dir():
source = source / SAMPLER_NAME
payload = torch.load(source, map_location="cpu", mmap=True, weights_only=False)
torch.save(slim_sampler_payload(payload), destination)
return destination
def as_condition_tensor(
condition: np.ndarray | torch.Tensor | list,
*,
device: str | torch.device,
) -> torch.Tensor:
if torch.is_tensor(condition):
t = condition.detach().float()
else:
t = torch.from_numpy(np.asarray(condition, dtype=np.float32))
if t.ndim == 1:
t = t.unsqueeze(0).unsqueeze(0) # [1,1,D]
elif t.ndim == 2:
t = t.unsqueeze(0) # [1,K,D]
elif t.ndim != 3:
raise ValueError(
f"condition embeddings must be [D], [K,D], or [B,K,D]; got {tuple(t.shape)}"
)
return t.to(device)
|