Spaces:
Build error
Build error
File size: 15,626 Bytes
b6f39f4 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 | """Sampling methods for diffusion models.
This module implements different sampling strategies for denoising diffusion models:
- DDPM: Original sampling from Ho et al. (2020)
- DDIM: Deterministic sampling from Song et al. (2021) - faster inference
"""
from abc import ABC, abstractmethod
import torch
from tqdm import tqdm
from ddpm.config import DiffusionConfig
from ddpm.models import ConditionalUNet
class DiffusionSampler(ABC):
"""Abstract base class for diffusion samplers.
Args:
config: Diffusion configuration
model: Conditional U-Net for denoising
betas: Noise schedule
alphas_cumprod: Cumulative product of alphas
"""
def __init__(
self,
config: DiffusionConfig,
model: ConditionalUNet,
betas: torch.Tensor,
alphas_cumprod: torch.Tensor,
):
self.config = config
self.model = model
self.device = config.device
self.betas = betas
self.alphas_cumprod = alphas_cumprod
self.alphas = 1.0 - betas
@abstractmethod
def sample(
self,
conditioning: torch.Tensor,
n_samples: int = 1,
require_grad: bool = False,
initial_noise: torch.Tensor | None = None,
show_progress: bool = True,
) -> torch.Tensor:
"""Generate samples given conditioning data.
Args:
conditioning: Conditioning data
n_samples: Number of samples to generate
require_grad: If True, enable gradient computation
initial_noise: Optional fixed initial noise tensor
show_progress: If True, show tqdm progress bar
Returns:
Generated samples
"""
pass
def _prepare_sampling(
self,
conditioning: torch.Tensor,
n_samples: int,
require_grad: bool,
initial_noise: torch.Tensor | None,
) -> tuple[torch.Tensor, torch.Tensor]:
"""Prepare conditioning and initial noise for sampling.
Handles model eval mode, conditioning expansion, device transfer,
gradient setup, and noise initialisation.
Args:
conditioning: Raw conditioning data
n_samples: Number of samples per conditioning
require_grad: Whether to enable gradient computation
initial_noise: Optional fixed initial noise
Returns:
Tuple of (prepared conditioning, initial noise x_t)
"""
if not require_grad:
self.model.eval()
# Ensure conditioning has batch dimension
if conditioning.dim() == 1:
conditioning = conditioning.unsqueeze(0).repeat(n_samples, 1)
elif conditioning.dim() == 2:
if self.config.n_conditioning_channels > 1:
conditioning = conditioning.unsqueeze(0).repeat(n_samples, 1, 1)
else:
conditioning = conditioning.repeat_interleave(n_samples, dim=0)
elif conditioning.dim() == 3:
conditioning = conditioning.repeat_interleave(n_samples, dim=0)
conditioning = conditioning.to(self.device)
if require_grad and not conditioning.requires_grad:
conditioning.requires_grad_(True)
if initial_noise is not None:
x_t = initial_noise.to(self.device)
else:
x_t = torch.randn(
conditioning.shape[0], self.config.trajectory_length, device=self.device
)
return conditioning, x_t
class DDPMSampler(DiffusionSampler):
"""DDPM sampling (Ho et al., 2020).
Uses the full reverse diffusion process with stochastic sampling.
Requires all T timesteps for generation.
"""
def __init__(
self,
config: DiffusionConfig,
model: ConditionalUNet,
betas: torch.Tensor,
alphas_cumprod: torch.Tensor,
):
super().__init__(config, model, betas, alphas_cumprod)
# Precompute posterior variance for efficiency
alphas_cumprod_prev = torch.nn.functional.pad(alphas_cumprod[:-1], (1, 0), value=1.0)
self.posterior_variance = betas * (1.0 - alphas_cumprod_prev) / (1.0 - alphas_cumprod)
def reverse_step(
self, x_t: torch.Tensor, t: torch.Tensor, conditioning: torch.Tensor, add_noise: bool = True
) -> torch.Tensor:
"""Single DDPM reverse diffusion step.
Args:
x_t: Noisy data at timestep t
t: Current timestep
conditioning: Conditioning data
add_noise: Whether to add noise (False for t=0)
Returns:
x_{t-1}: Less noisy data
"""
# Predict noise
if add_noise or x_t.requires_grad or conditioning.requires_grad:
predicted_noise = self.model(x_t, t, conditioning)
else:
with torch.no_grad():
predicted_noise = self.model(x_t, t, conditioning)
# Extract coefficients
alpha = self.alphas[t][:, None]
alpha_cumprod = self.alphas_cumprod[t][:, None]
beta = self.betas[t][:, None]
# Compute mean
mean = (x_t - beta * predicted_noise / torch.sqrt(1.0 - alpha_cumprod)) / torch.sqrt(alpha)
if add_noise and t[0] > 0:
# Add noise except for the last step
posterior_variance = self.posterior_variance[t][:, None]
noise = torch.randn_like(x_t)
x_t_minus_1 = mean + torch.sqrt(posterior_variance) * noise
else:
x_t_minus_1 = mean
return x_t_minus_1
def sample(
self,
conditioning: torch.Tensor,
n_samples: int = 1,
require_grad: bool = False,
initial_noise: torch.Tensor | None = None,
show_progress: bool = True,
) -> torch.Tensor:
"""Generate samples using DDPM sampling.
Args:
conditioning: Conditioning data of shape [trajectory_length] or [batch_size, trajectory_length]
n_samples: Number of samples to generate per conditioning
require_grad: If True, enable gradient computation
initial_noise: Optional fixed initial noise tensor. If provided, must have shape
[batch_size * n_samples, trajectory_length]. Useful for consistent
gradients in optimisation.
show_progress: If True, show tqdm progress bar. Disable for optimisation loops.
Returns:
Generated samples of shape [n_samples, trajectory_length]
"""
conditioning, x_t = self._prepare_sampling(
conditioning, n_samples, require_grad, initial_noise
)
# Reverse diffusion
timestep_iterator = range(self.config.n_timesteps - 1, -1, -1)
if show_progress:
timestep_iterator = tqdm(timestep_iterator, desc="DDPM Sampling")
for t_idx in timestep_iterator:
t = torch.full((conditioning.shape[0],), t_idx, dtype=torch.long).to(self.device)
x_t = self.reverse_step(x_t, t, conditioning, add_noise=(t_idx > 0))
return x_t
class DDIMSampler(DiffusionSampler):
"""DDIM sampling (Song et al., 2021).
Deterministic (or semi-deterministic) sampling that allows for faster inference
by skipping timesteps. Can generate samples in fewer steps than DDPM.
Args:
config: Diffusion configuration
model: Conditional U-Net for denoising
betas: Noise schedule
alphas_cumprod: Cumulative product of alphas
eta: Controls stochasticity (0 = deterministic, 1 = stochastic like DDPM)
num_inference_steps: Number of steps to use (can be less than training steps)
"""
def __init__(
self,
config: DiffusionConfig,
model: ConditionalUNet,
betas: torch.Tensor,
alphas_cumprod: torch.Tensor,
eta: float = 0.0,
num_inference_steps: int | None = None,
):
super().__init__(config, model, betas, alphas_cumprod)
self.eta = eta
# Use fewer steps for faster inference
self.num_inference_steps = num_inference_steps or config.n_timesteps
# Create timestep schedule (evenly spaced) while honouring the
# requested step count exactly. The schedule is stored in ascending
# order and reversed at sampling time.
if self.num_inference_steps < config.n_timesteps:
self.timesteps = (
torch.linspace(
config.n_timesteps - 1,
0,
steps=self.num_inference_steps,
)
.round()
.long()
.flip(0)
)
self.timesteps = torch.unique_consecutive(self.timesteps)
else:
self.timesteps = torch.arange(config.n_timesteps).long()
def reverse_step(
self, x_t: torch.Tensor, t: int, t_prev: int, conditioning: torch.Tensor
) -> torch.Tensor:
"""Single DDIM reverse step.
Uses the deterministic (or semi-deterministic) DDIM update rule:
x_{t-1} = sqrt(alpha_{t-1}) * pred_x0 + sqrt(1 - alpha_{t-1} - sigma_t^2) * epsilon + sigma_t * noise
where pred_x0 = (x_t - sqrt(1 - alpha_t) * epsilon) / sqrt(alpha_t)
Args:
x_t: Noisy data at timestep t
t: Current timestep index
t_prev: Previous timestep index (can skip timesteps)
conditioning: Conditioning data
Returns:
x_{t_prev}: Less noisy data at previous timestep
"""
# Predict noise
t_tensor = torch.full((x_t.shape[0],), t, dtype=torch.long, device=self.device)
if x_t.requires_grad or conditioning.requires_grad:
predicted_noise = self.model(x_t, t_tensor, conditioning)
else:
with torch.no_grad():
predicted_noise = self.model(x_t, t_tensor, conditioning)
# Get alpha values
alpha_t = self.alphas_cumprod[t]
alpha_t_prev = self.alphas_cumprod[t_prev] if t_prev >= 0 else torch.tensor(1.0)
# Predict x_0 from x_t and predicted noise
pred_x0 = (x_t - torch.sqrt(1.0 - alpha_t) * predicted_noise) / torch.sqrt(alpha_t)
# Compute variance (eta controls stochasticity)
sigma_t = self.eta * torch.sqrt(
(1.0 - alpha_t_prev) / (1.0 - alpha_t) * (1.0 - alpha_t / alpha_t_prev)
)
# Compute direction pointing to x_t
dir_xt = torch.sqrt(1.0 - alpha_t_prev - sigma_t**2) * predicted_noise
# Compute x_{t-1}
x_t_prev = torch.sqrt(alpha_t_prev) * pred_x0 + dir_xt
# Add noise if eta > 0 and not at the last step
if self.eta > 0 and t_prev >= 0:
noise = torch.randn_like(x_t)
x_t_prev = x_t_prev + sigma_t * noise
return x_t_prev
def sample(
self,
conditioning: torch.Tensor,
n_samples: int = 1,
require_grad: bool = False,
initial_noise: torch.Tensor | None = None,
show_progress: bool = True,
grad_steps: int | None = None,
grad_steps_spread: bool = False,
) -> torch.Tensor:
"""Generate samples using DDIM sampling.
Args:
conditioning: Conditioning data of shape [trajectory_length] or [batch_size, trajectory_length]
n_samples: Number of samples to generate per conditioning
require_grad: If True, enable gradient computation
initial_noise: Optional fixed initial noise tensor. If provided, must have shape
[batch_size * n_samples, trajectory_length]. Useful for consistent
gradients in optimisation.
show_progress: If True, show tqdm progress bar. Disable for optimisation loops.
grad_steps: If set, restricts autograd to a subset of denoising steps. By
default (``grad_steps_spread=False``) these are the *final* ``grad_steps``
steps (largest denoising effect, lowest noise). If ``grad_steps_spread=True``
the steps are chosen *linearly across the full chain*, capturing gradient
signal at the start, middle, and end of denoising.
grad_steps_spread: If ``True``, spread the ``grad_steps`` gradient-carrying
steps uniformly across the chain instead of concentrating them at the end.
Returns:
Generated samples of shape [n_samples, trajectory_length]
"""
conditioning, x_t = self._prepare_sampling(
conditioning, n_samples, require_grad, initial_noise
)
if x_t.shape[0] != conditioning.shape[0]:
raise ValueError(
f"initial_noise batch size ({x_t.shape[0]}) must match "
f"conditioning batch size ({conditioning.shape[0]})"
)
# Reverse diffusion with potentially fewer steps
timesteps_reversed = self.timesteps.flip(0)
n_total = len(timesteps_reversed)
# Determine which step indices carry gradients
if grad_steps is None:
grad_indices: set[int] | None = None
elif grad_steps_spread:
# Linearly spaced across the full chain (inclusive of first and last)
grad_indices = set(
int(round(i))
for i in torch.linspace(0, n_total - 1, min(grad_steps, n_total)).tolist()
)
else:
# Final grad_steps steps only
grad_indices = set(range(n_total - grad_steps, n_total))
if show_progress:
desc = f"DDIM Sampling ({self.num_inference_steps} steps)"
timestep_iterator = enumerate(tqdm(timesteps_reversed, desc=desc))
else:
timestep_iterator = enumerate(timesteps_reversed)
for i, t_idx in timestep_iterator:
t = int(t_idx.item())
# Get previous timestep (or -1 for the last step)
t_prev = (
int(timesteps_reversed[i + 1].item()) if i + 1 < len(timesteps_reversed) else -1
)
# Optionally restrict autograd graph to selected steps
if grad_indices is not None and i not in grad_indices:
with torch.no_grad():
x_t = self.reverse_step(x_t.detach(), t, t_prev, conditioning.detach())
else:
x_t = self.reverse_step(x_t, t, t_prev, conditioning)
return x_t
def create_sampler(
sampler_type: str,
config: DiffusionConfig,
model: ConditionalUNet,
betas: torch.Tensor,
alphas_cumprod: torch.Tensor,
**kwargs,
) -> DiffusionSampler:
"""Factory function to create a sampler.
Args:
sampler_type: Type of sampler ('ddpm' or 'ddim')
config: Diffusion configuration
model: Conditional U-Net
betas: Noise schedule
alphas_cumprod: Cumulative product of alphas
**kwargs: Additional arguments for specific samplers
- For DDIM: eta (float), num_inference_steps (int)
Returns:
Sampler instance
"""
sampler_type = sampler_type.lower()
if sampler_type == "ddpm":
return DDPMSampler(config, model, betas, alphas_cumprod)
elif sampler_type == "ddim":
return DDIMSampler(
config,
model,
betas,
alphas_cumprod,
eta=kwargs.get("eta", 0.0),
num_inference_steps=kwargs.get("num_inference_steps", None),
)
else:
raise ValueError(f"Unknown sampler type: {sampler_type}. Choose 'ddpm' or 'ddim'.")
|