Instructions to use BiliSakura/ADM-diffusers with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use BiliSakura/ADM-diffusers with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("BiliSakura/ADM-diffusers", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
- Local Apps
- Draw Things
- DiffusionBee
Delete ADM-G-256/scheduler/scheduling_adm.py
Browse files
ADM-G-256/scheduler/scheduling_adm.py
DELETED
|
@@ -1,590 +0,0 @@
|
|
| 1 |
-
# Copyright 2026 The HuggingFace Team. All rights reserved.
|
| 2 |
-
#
|
| 3 |
-
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 4 |
-
# you may not use this file except in compliance with the License.
|
| 5 |
-
|
| 6 |
-
import enum
|
| 7 |
-
import math
|
| 8 |
-
from dataclasses import dataclass
|
| 9 |
-
from typing import Optional, Tuple, Union
|
| 10 |
-
|
| 11 |
-
import numpy as np
|
| 12 |
-
import torch
|
| 13 |
-
|
| 14 |
-
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
| 15 |
-
from diffusers.schedulers.scheduling_utils import SchedulerMixin
|
| 16 |
-
from diffusers.utils import BaseOutput
|
| 17 |
-
|
| 18 |
-
try:
|
| 19 |
-
from diffusers.utils.torch_utils import randn_tensor
|
| 20 |
-
except ImportError: # pragma: no cover
|
| 21 |
-
def randn_tensor(shape, generator=None, device=None, dtype=None):
|
| 22 |
-
return torch.randn(shape, generator=generator, device=device, dtype=dtype)
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
# ---------------------------------------------------------------------------
|
| 26 |
-
# Internal diffusion math (OpenAI ADM / improved-diffusion)
|
| 27 |
-
# ---------------------------------------------------------------------------
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
def _randn_like(tensor: torch.Tensor, generator: Optional[torch.Generator] = None) -> torch.Tensor:
|
| 31 |
-
return randn_tensor(tensor.shape, generator=generator, device=tensor.device, dtype=tensor.dtype)
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
def _extract_into_tensor(arr, timesteps, broadcast_shape):
|
| 35 |
-
res = torch.from_numpy(arr).to(device=timesteps.device)[timesteps].float()
|
| 36 |
-
while len(res.shape) < len(broadcast_shape):
|
| 37 |
-
res = res[..., None]
|
| 38 |
-
return res.expand(broadcast_shape)
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
def _get_named_beta_schedule(schedule_name: str, num_diffusion_timesteps: int):
|
| 42 |
-
if schedule_name == "linear":
|
| 43 |
-
scale = 1000 / num_diffusion_timesteps
|
| 44 |
-
return np.linspace(scale * 0.0001, scale * 0.02, num_diffusion_timesteps, dtype=np.float64)
|
| 45 |
-
if schedule_name == "cosine":
|
| 46 |
-
return _betas_for_alpha_bar(
|
| 47 |
-
num_diffusion_timesteps,
|
| 48 |
-
lambda t: math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2,
|
| 49 |
-
)
|
| 50 |
-
raise NotImplementedError(f"unknown beta schedule: {schedule_name}")
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
def _betas_for_alpha_bar(num_diffusion_timesteps: int, alpha_bar, max_beta: float = 0.999):
|
| 54 |
-
betas = []
|
| 55 |
-
for i in range(num_diffusion_timesteps):
|
| 56 |
-
t1 = i / num_diffusion_timesteps
|
| 57 |
-
t2 = (i + 1) / num_diffusion_timesteps
|
| 58 |
-
betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
|
| 59 |
-
return np.array(betas)
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
def _space_timesteps(num_timesteps: int, section_counts):
|
| 63 |
-
if isinstance(section_counts, str):
|
| 64 |
-
if section_counts.startswith("ddim"):
|
| 65 |
-
desired_count = int(section_counts[len("ddim") :])
|
| 66 |
-
for i in range(1, num_timesteps):
|
| 67 |
-
if len(range(0, num_timesteps, i)) == desired_count:
|
| 68 |
-
return set(range(0, num_timesteps, i))
|
| 69 |
-
raise ValueError(f"cannot create exactly {num_timesteps} steps with an integer stride")
|
| 70 |
-
section_counts = [int(x) for x in section_counts.split(",")]
|
| 71 |
-
|
| 72 |
-
size_per = num_timesteps // len(section_counts)
|
| 73 |
-
extra = num_timesteps % len(section_counts)
|
| 74 |
-
start_idx = 0
|
| 75 |
-
all_steps = []
|
| 76 |
-
for i, section_count in enumerate(section_counts):
|
| 77 |
-
size = size_per + (1 if i < extra else 0)
|
| 78 |
-
if size < section_count:
|
| 79 |
-
raise ValueError(f"cannot divide section of {size} steps into {section_count}")
|
| 80 |
-
frac_stride = 1 if section_count <= 1 else (size - 1) / (section_count - 1)
|
| 81 |
-
cur_idx = 0.0
|
| 82 |
-
for _ in range(section_count):
|
| 83 |
-
all_steps.append(start_idx + round(cur_idx))
|
| 84 |
-
cur_idx += frac_stride
|
| 85 |
-
start_idx += size
|
| 86 |
-
return set(all_steps)
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
class _ModelMeanType(enum.Enum):
|
| 90 |
-
PREVIOUS_X = enum.auto()
|
| 91 |
-
START_X = enum.auto()
|
| 92 |
-
EPSILON = enum.auto()
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
class _ModelVarType(enum.Enum):
|
| 96 |
-
LEARNED = enum.auto()
|
| 97 |
-
FIXED_SMALL = enum.auto()
|
| 98 |
-
FIXED_LARGE = enum.auto()
|
| 99 |
-
LEARNED_RANGE = enum.auto()
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
class _GaussianDiffusion:
|
| 103 |
-
def __init__(self, *, betas, model_mean_type, model_var_type, rescale_timesteps: bool = False):
|
| 104 |
-
self.model_mean_type = model_mean_type
|
| 105 |
-
self.model_var_type = model_var_type
|
| 106 |
-
self.rescale_timesteps = rescale_timesteps
|
| 107 |
-
betas = np.array(betas, dtype=np.float64)
|
| 108 |
-
self.betas = betas
|
| 109 |
-
self.num_timesteps = int(betas.shape[0])
|
| 110 |
-
|
| 111 |
-
alphas = 1.0 - betas
|
| 112 |
-
self.alphas_cumprod = np.cumprod(alphas, axis=0)
|
| 113 |
-
self.alphas_cumprod_prev = np.append(1.0, self.alphas_cumprod[:-1])
|
| 114 |
-
self.sqrt_recip_alphas_cumprod = np.sqrt(1.0 / self.alphas_cumprod)
|
| 115 |
-
self.sqrt_recipm1_alphas_cumprod = np.sqrt(1.0 / self.alphas_cumprod - 1)
|
| 116 |
-
self.posterior_variance = betas * (1.0 - self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod)
|
| 117 |
-
self.posterior_log_variance_clipped = np.log(np.append(self.posterior_variance[1], self.posterior_variance[1:]))
|
| 118 |
-
self.posterior_mean_coef1 = betas * np.sqrt(self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod)
|
| 119 |
-
self.posterior_mean_coef2 = (1.0 - self.alphas_cumprod_prev) * np.sqrt(alphas) / (1.0 - self.alphas_cumprod)
|
| 120 |
-
|
| 121 |
-
def _predict_xstart_from_eps(self, x_t, t, eps):
|
| 122 |
-
return _extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - _extract_into_tensor(
|
| 123 |
-
self.sqrt_recipm1_alphas_cumprod, t, x_t.shape
|
| 124 |
-
) * eps
|
| 125 |
-
|
| 126 |
-
def _predict_eps_from_xstart(self, x_t, t, pred_xstart):
|
| 127 |
-
return (
|
| 128 |
-
_extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart
|
| 129 |
-
) / _extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
|
| 130 |
-
|
| 131 |
-
def _predict_xstart_from_xprev(self, x_t, t, xprev):
|
| 132 |
-
return _extract_into_tensor(1.0 / self.posterior_mean_coef1, t, x_t.shape) * xprev - _extract_into_tensor(
|
| 133 |
-
self.posterior_mean_coef2 / self.posterior_mean_coef1, t, x_t.shape
|
| 134 |
-
) * x_t
|
| 135 |
-
|
| 136 |
-
def q_posterior_mean_variance(self, x_start, x_t, t):
|
| 137 |
-
posterior_mean = _extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start + _extract_into_tensor(
|
| 138 |
-
self.posterior_mean_coef2, t, x_t.shape
|
| 139 |
-
) * x_t
|
| 140 |
-
posterior_variance = _extract_into_tensor(self.posterior_variance, t, x_t.shape)
|
| 141 |
-
posterior_log_variance_clipped = _extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape)
|
| 142 |
-
return posterior_mean, posterior_variance, posterior_log_variance_clipped
|
| 143 |
-
|
| 144 |
-
def p_mean_variance_from_output(
|
| 145 |
-
self,
|
| 146 |
-
model_output: torch.Tensor,
|
| 147 |
-
x: torch.Tensor,
|
| 148 |
-
t: torch.Tensor,
|
| 149 |
-
clip_denoised: bool = True,
|
| 150 |
-
):
|
| 151 |
-
_, c = x.shape[:2]
|
| 152 |
-
|
| 153 |
-
if self.model_var_type == _ModelVarType.LEARNED_RANGE:
|
| 154 |
-
model_output, model_var_values = torch.split(model_output, c, dim=1)
|
| 155 |
-
min_log = _extract_into_tensor(self.posterior_log_variance_clipped, t, x.shape)
|
| 156 |
-
max_log = _extract_into_tensor(np.log(self.betas), t, x.shape)
|
| 157 |
-
frac = (model_var_values + 1) / 2
|
| 158 |
-
model_log_variance = frac * max_log + (1 - frac) * min_log
|
| 159 |
-
model_variance = torch.exp(model_log_variance)
|
| 160 |
-
else:
|
| 161 |
-
model_variance, model_log_variance = {
|
| 162 |
-
_ModelVarType.FIXED_LARGE: (
|
| 163 |
-
np.append(self.posterior_variance[1], self.betas[1:]),
|
| 164 |
-
np.log(np.append(self.posterior_variance[1], self.betas[1:])),
|
| 165 |
-
),
|
| 166 |
-
_ModelVarType.FIXED_SMALL: (self.posterior_variance, self.posterior_log_variance_clipped),
|
| 167 |
-
}[self.model_var_type]
|
| 168 |
-
model_variance = _extract_into_tensor(model_variance, t, x.shape)
|
| 169 |
-
model_log_variance = _extract_into_tensor(model_log_variance, t, x.shape)
|
| 170 |
-
|
| 171 |
-
if self.model_mean_type == _ModelMeanType.START_X:
|
| 172 |
-
pred_xstart = model_output
|
| 173 |
-
elif self.model_mean_type == _ModelMeanType.EPSILON:
|
| 174 |
-
pred_xstart = self._predict_xstart_from_eps(x_t=x, t=t, eps=model_output)
|
| 175 |
-
else:
|
| 176 |
-
pred_xstart = self._predict_xstart_from_xprev(x_t=x, t=t, xprev=model_output)
|
| 177 |
-
if clip_denoised:
|
| 178 |
-
pred_xstart = pred_xstart.clamp(-1, 1)
|
| 179 |
-
model_mean, _, _ = self.q_posterior_mean_variance(x_start=pred_xstart, x_t=x, t=t)
|
| 180 |
-
return {"mean": model_mean, "variance": model_variance, "log_variance": model_log_variance, "pred_xstart": pred_xstart}
|
| 181 |
-
|
| 182 |
-
def p_mean_variance(self, model, x, t, clip_denoised: bool = True, model_kwargs=None):
|
| 183 |
-
model_kwargs = {} if model_kwargs is None else model_kwargs
|
| 184 |
-
if self.rescale_timesteps:
|
| 185 |
-
ts = t.float() * (1000.0 / self.num_timesteps)
|
| 186 |
-
else:
|
| 187 |
-
ts = t
|
| 188 |
-
model_output = model(x, ts, **model_kwargs)
|
| 189 |
-
return self.p_mean_variance_from_output(model_output, x, t, clip_denoised=clip_denoised)
|
| 190 |
-
|
| 191 |
-
def condition_mean(self, cond_grad: torch.Tensor, p_mean_var: dict, x: torch.Tensor) -> torch.Tensor:
|
| 192 |
-
"""Apply classifier guidance to the reverse-process mean (Sohl-Dickstein et al., 2015)."""
|
| 193 |
-
del x
|
| 194 |
-
return p_mean_var["mean"].float() + p_mean_var["variance"] * cond_grad.float()
|
| 195 |
-
|
| 196 |
-
def p_sample_from_output(
|
| 197 |
-
self,
|
| 198 |
-
model_output: torch.Tensor,
|
| 199 |
-
x: torch.Tensor,
|
| 200 |
-
t: torch.Tensor,
|
| 201 |
-
clip_denoised: bool = True,
|
| 202 |
-
generator: Optional[torch.Generator] = None,
|
| 203 |
-
cond_grad: Optional[torch.Tensor] = None,
|
| 204 |
-
):
|
| 205 |
-
out = self.p_mean_variance_from_output(model_output, x, t, clip_denoised=clip_denoised)
|
| 206 |
-
if cond_grad is not None:
|
| 207 |
-
out["mean"] = self.condition_mean(cond_grad, out, x)
|
| 208 |
-
noise = _randn_like(x, generator=generator)
|
| 209 |
-
nonzero_mask = (t != 0).float().view(-1, *([1] * (len(x.shape) - 1)))
|
| 210 |
-
sample = out["mean"] + nonzero_mask * torch.exp(0.5 * out["log_variance"]) * noise
|
| 211 |
-
return {"sample": sample, "pred_xstart": out["pred_xstart"]}
|
| 212 |
-
|
| 213 |
-
def p_sample(self, model, x, t, clip_denoised=True, model_kwargs=None, generator: Optional[torch.Generator] = None):
|
| 214 |
-
out = self.p_mean_variance(model, x, t, clip_denoised=clip_denoised, model_kwargs=model_kwargs)
|
| 215 |
-
noise = _randn_like(x, generator=generator)
|
| 216 |
-
nonzero_mask = (t != 0).float().view(-1, *([1] * (len(x.shape) - 1)))
|
| 217 |
-
sample = out["mean"] + nonzero_mask * torch.exp(0.5 * out["log_variance"]) * noise
|
| 218 |
-
return {"sample": sample, "pred_xstart": out["pred_xstart"]}
|
| 219 |
-
|
| 220 |
-
def p_sample_loop(self, model, shape, noise=None, clip_denoised=True, model_kwargs=None, device=None, progress=False):
|
| 221 |
-
final = None
|
| 222 |
-
for sample in self.p_sample_loop_progressive(
|
| 223 |
-
model, shape, noise=noise, clip_denoised=clip_denoised, model_kwargs=model_kwargs, device=device, progress=progress
|
| 224 |
-
):
|
| 225 |
-
final = sample
|
| 226 |
-
return final["sample"]
|
| 227 |
-
|
| 228 |
-
def p_sample_loop_progressive(self, model, shape, noise=None, clip_denoised=True, model_kwargs=None, device=None, progress=False):
|
| 229 |
-
if device is None:
|
| 230 |
-
device = next(model.parameters()).device
|
| 231 |
-
img = noise if noise is not None else torch.randn(*shape, device=device)
|
| 232 |
-
indices = list(range(self.num_timesteps))[::-1]
|
| 233 |
-
if progress:
|
| 234 |
-
from tqdm.auto import tqdm
|
| 235 |
-
|
| 236 |
-
indices = tqdm(indices)
|
| 237 |
-
for i in indices:
|
| 238 |
-
t = torch.tensor([i] * shape[0], device=device)
|
| 239 |
-
with torch.no_grad():
|
| 240 |
-
out = self.p_sample(model, img, t, clip_denoised=clip_denoised, model_kwargs=model_kwargs)
|
| 241 |
-
yield out
|
| 242 |
-
img = out["sample"]
|
| 243 |
-
|
| 244 |
-
def condition_score(self, cond_grad: torch.Tensor, p_mean_var: dict, x: torch.Tensor, t: torch.Tensor) -> dict:
|
| 245 |
-
"""Apply classifier guidance to the score (Song et al., 2020) for DDIM."""
|
| 246 |
-
alpha_bar = _extract_into_tensor(self.alphas_cumprod, t, x.shape)
|
| 247 |
-
eps = self._predict_eps_from_xstart(x, t, p_mean_var["pred_xstart"])
|
| 248 |
-
eps = eps - (1 - alpha_bar).sqrt() * cond_grad
|
| 249 |
-
out = dict(p_mean_var)
|
| 250 |
-
out["pred_xstart"] = self._predict_xstart_from_eps(x_t=x, t=t, eps=eps)
|
| 251 |
-
out["mean"], _, _ = self.q_posterior_mean_variance(x_start=out["pred_xstart"], x_t=x, t=t)
|
| 252 |
-
return out
|
| 253 |
-
|
| 254 |
-
def ddim_sample_from_output(
|
| 255 |
-
self,
|
| 256 |
-
model_output: torch.Tensor,
|
| 257 |
-
x: torch.Tensor,
|
| 258 |
-
t: torch.Tensor,
|
| 259 |
-
clip_denoised: bool = True,
|
| 260 |
-
eta: float = 0.0,
|
| 261 |
-
generator: Optional[torch.Generator] = None,
|
| 262 |
-
cond_grad: Optional[torch.Tensor] = None,
|
| 263 |
-
):
|
| 264 |
-
out = self.p_mean_variance_from_output(model_output, x, t, clip_denoised=clip_denoised)
|
| 265 |
-
if cond_grad is not None:
|
| 266 |
-
out = self.condition_score(cond_grad, out, x, t)
|
| 267 |
-
pred_xstart = out["pred_xstart"]
|
| 268 |
-
eps = self._predict_eps_from_xstart(x, t, pred_xstart)
|
| 269 |
-
alpha_bar = _extract_into_tensor(self.alphas_cumprod, t, x.shape)
|
| 270 |
-
alpha_bar_prev = _extract_into_tensor(self.alphas_cumprod_prev, t, x.shape)
|
| 271 |
-
sigma = eta * torch.sqrt((1 - alpha_bar_prev) / (1 - alpha_bar)) * torch.sqrt(1 - alpha_bar / alpha_bar_prev)
|
| 272 |
-
noise = _randn_like(x, generator=generator)
|
| 273 |
-
mean_pred = pred_xstart * torch.sqrt(alpha_bar_prev) + torch.sqrt(1 - alpha_bar_prev - sigma**2) * eps
|
| 274 |
-
nonzero_mask = (t != 0).float().view(-1, *([1] * (len(x.shape) - 1)))
|
| 275 |
-
sample = mean_pred + nonzero_mask * sigma * noise
|
| 276 |
-
return {"sample": sample, "pred_xstart": pred_xstart}
|
| 277 |
-
|
| 278 |
-
def ddim_sample(
|
| 279 |
-
self,
|
| 280 |
-
model,
|
| 281 |
-
x,
|
| 282 |
-
t,
|
| 283 |
-
clip_denoised=True,
|
| 284 |
-
model_kwargs=None,
|
| 285 |
-
eta=0.0,
|
| 286 |
-
generator: Optional[torch.Generator] = None,
|
| 287 |
-
):
|
| 288 |
-
model_kwargs = {} if model_kwargs is None else model_kwargs
|
| 289 |
-
if self.rescale_timesteps:
|
| 290 |
-
ts = t.float() * (1000.0 / self.num_timesteps)
|
| 291 |
-
else:
|
| 292 |
-
ts = t
|
| 293 |
-
model_output = model(x, ts, **model_kwargs)
|
| 294 |
-
return self.ddim_sample_from_output(
|
| 295 |
-
model_output, x, t, clip_denoised=clip_denoised, eta=eta, generator=generator
|
| 296 |
-
)
|
| 297 |
-
|
| 298 |
-
|
| 299 |
-
class _WrappedModel:
|
| 300 |
-
def __init__(self, model, timestep_map, rescale_timesteps, original_num_steps):
|
| 301 |
-
self.model = model
|
| 302 |
-
self.timestep_map = timestep_map
|
| 303 |
-
self.rescale_timesteps = rescale_timesteps
|
| 304 |
-
self.original_num_steps = original_num_steps
|
| 305 |
-
|
| 306 |
-
def __call__(self, x, ts, **kwargs):
|
| 307 |
-
map_tensor = torch.tensor(self.timestep_map, device=ts.device, dtype=ts.dtype)
|
| 308 |
-
new_ts = map_tensor[ts]
|
| 309 |
-
if self.rescale_timesteps:
|
| 310 |
-
new_ts = new_ts.float() * (1000.0 / self.original_num_steps)
|
| 311 |
-
return self.model(x, new_ts, **kwargs)
|
| 312 |
-
|
| 313 |
-
|
| 314 |
-
class _SpacedDiffusion(_GaussianDiffusion):
|
| 315 |
-
def __init__(self, use_timesteps, **kwargs):
|
| 316 |
-
self.use_timesteps = set(use_timesteps)
|
| 317 |
-
self.timestep_map = []
|
| 318 |
-
self.original_num_steps = len(kwargs["betas"])
|
| 319 |
-
base_diffusion = _GaussianDiffusion(**kwargs)
|
| 320 |
-
last_alpha_cumprod = 1.0
|
| 321 |
-
new_betas = []
|
| 322 |
-
for i, alpha_cumprod in enumerate(base_diffusion.alphas_cumprod):
|
| 323 |
-
if i in self.use_timesteps:
|
| 324 |
-
new_betas.append(1 - alpha_cumprod / last_alpha_cumprod)
|
| 325 |
-
last_alpha_cumprod = alpha_cumprod
|
| 326 |
-
self.timestep_map.append(i)
|
| 327 |
-
kwargs["betas"] = np.array(new_betas)
|
| 328 |
-
super().__init__(**kwargs)
|
| 329 |
-
|
| 330 |
-
def p_mean_variance(self, model, *args, **kwargs):
|
| 331 |
-
return super().p_mean_variance(self._wrap_model(model), *args, **kwargs)
|
| 332 |
-
|
| 333 |
-
def _wrap_model(self, model):
|
| 334 |
-
if isinstance(model, _WrappedModel):
|
| 335 |
-
return model
|
| 336 |
-
return _WrappedModel(model, self.timestep_map, self.rescale_timesteps, self.original_num_steps)
|
| 337 |
-
|
| 338 |
-
|
| 339 |
-
def _create_spaced_diffusion(
|
| 340 |
-
*,
|
| 341 |
-
steps: int = 1000,
|
| 342 |
-
learn_sigma: bool = False,
|
| 343 |
-
sigma_small: bool = False,
|
| 344 |
-
noise_schedule: str = "linear",
|
| 345 |
-
predict_xstart: bool = False,
|
| 346 |
-
rescale_timesteps: bool = False,
|
| 347 |
-
timestep_respacing: str = "",
|
| 348 |
-
) -> _SpacedDiffusion:
|
| 349 |
-
betas = _get_named_beta_schedule(noise_schedule, steps)
|
| 350 |
-
if not timestep_respacing:
|
| 351 |
-
timestep_respacing = [steps]
|
| 352 |
-
return _SpacedDiffusion(
|
| 353 |
-
use_timesteps=_space_timesteps(steps, timestep_respacing),
|
| 354 |
-
betas=betas,
|
| 355 |
-
model_mean_type=_ModelMeanType.EPSILON if not predict_xstart else _ModelMeanType.START_X,
|
| 356 |
-
model_var_type=(_ModelVarType.FIXED_LARGE if not sigma_small else _ModelVarType.FIXED_SMALL)
|
| 357 |
-
if not learn_sigma
|
| 358 |
-
else _ModelVarType.LEARNED_RANGE,
|
| 359 |
-
rescale_timesteps=rescale_timesteps,
|
| 360 |
-
)
|
| 361 |
-
|
| 362 |
-
|
| 363 |
-
# ---------------------------------------------------------------------------
|
| 364 |
-
# Public Diffusers scheduler API
|
| 365 |
-
# ---------------------------------------------------------------------------
|
| 366 |
-
|
| 367 |
-
|
| 368 |
-
@dataclass
|
| 369 |
-
class ADMSchedulerOutput(BaseOutput):
|
| 370 |
-
"""
|
| 371 |
-
Output class for the ADM scheduler's `step` function.
|
| 372 |
-
|
| 373 |
-
Args:
|
| 374 |
-
prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)`):
|
| 375 |
-
Computed sample `(x_{t-1})` of the previous timestep. `prev_sample` should be used as the next model input.
|
| 376 |
-
pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)`, *optional*):
|
| 377 |
-
The predicted denoised sample `(x_{0})` based on the model output.
|
| 378 |
-
"""
|
| 379 |
-
|
| 380 |
-
prev_sample: torch.FloatTensor
|
| 381 |
-
pred_original_sample: Optional[torch.FloatTensor] = None
|
| 382 |
-
|
| 383 |
-
|
| 384 |
-
class ADMScheduler(SchedulerMixin, ConfigMixin):
|
| 385 |
-
"""
|
| 386 |
-
DDPM / DDIM scheduler for ADM (Ablated Diffusion Model) with OpenAI-style Gaussian diffusion.
|
| 387 |
-
|
| 388 |
-
This scheduler implements spaced diffusion used by ADM checkpoints. Call `set_timesteps` before inference, then
|
| 389 |
-
alternate UNet forward passes with `step`.
|
| 390 |
-
"""
|
| 391 |
-
|
| 392 |
-
config_name = "scheduler_config.json"
|
| 393 |
-
order = 1
|
| 394 |
-
|
| 395 |
-
@register_to_config
|
| 396 |
-
def __init__(
|
| 397 |
-
self,
|
| 398 |
-
steps: int = 1000,
|
| 399 |
-
learn_sigma: bool = False,
|
| 400 |
-
sigma_small: bool = False,
|
| 401 |
-
noise_schedule: str = "linear",
|
| 402 |
-
predict_xstart: bool = False,
|
| 403 |
-
rescale_timesteps: bool = False,
|
| 404 |
-
timestep_respacing: str = "",
|
| 405 |
-
):
|
| 406 |
-
self.timesteps = None
|
| 407 |
-
self.num_inference_steps = None
|
| 408 |
-
self._diffusion: Optional[_SpacedDiffusion] = None
|
| 409 |
-
self._use_ddim = False
|
| 410 |
-
self._eta = 0.0
|
| 411 |
-
|
| 412 |
-
def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor:
|
| 413 |
-
"""
|
| 414 |
-
Ensures interchangeability with schedulers that scale the denoising model input depending on the timestep.
|
| 415 |
-
|
| 416 |
-
Args:
|
| 417 |
-
sample (`torch.Tensor`):
|
| 418 |
-
The input sample.
|
| 419 |
-
timestep (`int`, *optional*):
|
| 420 |
-
The current timestep in the diffusion chain.
|
| 421 |
-
|
| 422 |
-
Returns:
|
| 423 |
-
`torch.Tensor`:
|
| 424 |
-
The (unchanged) input sample.
|
| 425 |
-
"""
|
| 426 |
-
del timestep
|
| 427 |
-
return sample
|
| 428 |
-
|
| 429 |
-
def set_timesteps(
|
| 430 |
-
self,
|
| 431 |
-
num_inference_steps: int,
|
| 432 |
-
device: Optional[Union[str, torch.device]] = None,
|
| 433 |
-
use_ddim: bool = False,
|
| 434 |
-
timestep_respacing: Optional[str] = None,
|
| 435 |
-
) -> torch.Tensor:
|
| 436 |
-
"""
|
| 437 |
-
Sets the discrete timesteps used for the diffusion chain (to be run before inference).
|
| 438 |
-
|
| 439 |
-
Args:
|
| 440 |
-
num_inference_steps (`int`):
|
| 441 |
-
The number of diffusion steps used when generating samples with a pre-trained model.
|
| 442 |
-
device (`str` or `torch.device`, *optional*):
|
| 443 |
-
The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
|
| 444 |
-
use_ddim (`bool`, *optional*, defaults to `False`):
|
| 445 |
-
Whether to use DDIM sampling instead of DDPM.
|
| 446 |
-
timestep_respacing (`str`, *optional*):
|
| 447 |
-
Override for the respacing string. If `None`, respacing is derived from `num_inference_steps`.
|
| 448 |
-
|
| 449 |
-
Returns:
|
| 450 |
-
`torch.Tensor`:
|
| 451 |
-
Timestep indices used during denoising, in descending order.
|
| 452 |
-
"""
|
| 453 |
-
if timestep_respacing is None:
|
| 454 |
-
timestep_respacing = f"ddim{num_inference_steps}" if use_ddim else str(num_inference_steps)
|
| 455 |
-
|
| 456 |
-
self._diffusion = _create_spaced_diffusion(
|
| 457 |
-
steps=self.config.steps,
|
| 458 |
-
learn_sigma=self.config.learn_sigma,
|
| 459 |
-
sigma_small=self.config.sigma_small,
|
| 460 |
-
noise_schedule=self.config.noise_schedule,
|
| 461 |
-
predict_xstart=self.config.predict_xstart,
|
| 462 |
-
rescale_timesteps=self.config.rescale_timesteps,
|
| 463 |
-
timestep_respacing=timestep_respacing,
|
| 464 |
-
)
|
| 465 |
-
self._use_ddim = use_ddim
|
| 466 |
-
self.num_inference_steps = num_inference_steps
|
| 467 |
-
|
| 468 |
-
indices = list(range(self._diffusion.num_timesteps))[::-1]
|
| 469 |
-
timesteps = torch.tensor(indices, dtype=torch.long)
|
| 470 |
-
if device is not None:
|
| 471 |
-
timesteps = timesteps.to(device)
|
| 472 |
-
self.timesteps = timesteps
|
| 473 |
-
return self.timesteps
|
| 474 |
-
|
| 475 |
-
def scale_timesteps_for_model(self, timestep: torch.Tensor) -> torch.Tensor:
|
| 476 |
-
"""
|
| 477 |
-
Map respaced scheduler indices to the timestep embeddings expected by the ADM UNet.
|
| 478 |
-
|
| 479 |
-
Args:
|
| 480 |
-
timestep (`torch.Tensor`):
|
| 481 |
-
Current scheduler timestep indices of shape `(batch_size,)`.
|
| 482 |
-
|
| 483 |
-
Returns:
|
| 484 |
-
`torch.Tensor`:
|
| 485 |
-
Timesteps to pass to the UNet forward pass.
|
| 486 |
-
"""
|
| 487 |
-
if self._diffusion is None:
|
| 488 |
-
raise ValueError("Call `set_timesteps` before running the scheduler.")
|
| 489 |
-
|
| 490 |
-
map_tensor = torch.tensor(self._diffusion.timestep_map, device=timestep.device, dtype=timestep.dtype)
|
| 491 |
-
model_timesteps = map_tensor[timestep]
|
| 492 |
-
if self._diffusion.rescale_timesteps:
|
| 493 |
-
model_timesteps = model_timesteps.float() * (1000.0 / self._diffusion.original_num_steps)
|
| 494 |
-
return model_timesteps
|
| 495 |
-
|
| 496 |
-
def step(
|
| 497 |
-
self,
|
| 498 |
-
model_output: torch.Tensor,
|
| 499 |
-
timestep: Union[int, torch.Tensor],
|
| 500 |
-
sample: torch.Tensor,
|
| 501 |
-
generator: Optional[torch.Generator] = None,
|
| 502 |
-
return_dict: bool = True,
|
| 503 |
-
clip_denoised: bool = True,
|
| 504 |
-
eta: Optional[float] = None,
|
| 505 |
-
cond_grad: Optional[torch.Tensor] = None,
|
| 506 |
-
) -> Union[ADMSchedulerOutput, Tuple[torch.Tensor, ...]]:
|
| 507 |
-
"""
|
| 508 |
-
Predict the sample at the previous timestep from the model output.
|
| 509 |
-
|
| 510 |
-
Args:
|
| 511 |
-
model_output (`torch.Tensor`):
|
| 512 |
-
The direct output from the ADM UNet.
|
| 513 |
-
timestep (`int` or `torch.Tensor`):
|
| 514 |
-
The current discrete timestep index in the respaced diffusion chain.
|
| 515 |
-
sample (`torch.Tensor`):
|
| 516 |
-
A current instance of a sample created by the diffusion process.
|
| 517 |
-
generator (`torch.Generator`, *optional*):
|
| 518 |
-
A random number generator for the sampling noise.
|
| 519 |
-
return_dict (`bool`, *optional*, defaults to `True`):
|
| 520 |
-
Whether or not to return an [`ADMSchedulerOutput`] instead of a plain tuple.
|
| 521 |
-
clip_denoised (`bool`, *optional*, defaults to `True`):
|
| 522 |
-
Whether to clamp the predicted `x_0` to `[-1, 1]`.
|
| 523 |
-
eta (`float`, *optional*):
|
| 524 |
-
DDIM stochasticity parameter. Only used when `use_ddim=True` was passed to `set_timesteps`.
|
| 525 |
-
cond_grad (`torch.Tensor`, *optional*):
|
| 526 |
-
Classifier guidance gradient for ADM-G (`classifier_scale * grad log p(y|x_t)`).
|
| 527 |
-
|
| 528 |
-
Returns:
|
| 529 |
-
[`ADMSchedulerOutput`] or `tuple`:
|
| 530 |
-
If `return_dict` is `True`, an [`ADMSchedulerOutput`] is returned, otherwise a tuple is returned where
|
| 531 |
-
the first element is the previous sample.
|
| 532 |
-
"""
|
| 533 |
-
if self._diffusion is None:
|
| 534 |
-
raise ValueError("Call `set_timesteps` before `step`.")
|
| 535 |
-
|
| 536 |
-
if not torch.is_tensor(timestep):
|
| 537 |
-
timestep = torch.tensor([timestep], device=sample.device, dtype=torch.long)
|
| 538 |
-
elif timestep.ndim == 0:
|
| 539 |
-
timestep = timestep.reshape(1).to(device=sample.device, dtype=torch.long)
|
| 540 |
-
else:
|
| 541 |
-
timestep = timestep.to(device=sample.device, dtype=torch.long)
|
| 542 |
-
|
| 543 |
-
ddim_eta = self._eta if eta is None else eta
|
| 544 |
-
|
| 545 |
-
if self._use_ddim:
|
| 546 |
-
out = self._diffusion.ddim_sample_from_output(
|
| 547 |
-
model_output,
|
| 548 |
-
sample,
|
| 549 |
-
timestep,
|
| 550 |
-
clip_denoised=clip_denoised,
|
| 551 |
-
eta=ddim_eta,
|
| 552 |
-
generator=generator,
|
| 553 |
-
cond_grad=cond_grad,
|
| 554 |
-
)
|
| 555 |
-
else:
|
| 556 |
-
out = self._diffusion.p_sample_from_output(
|
| 557 |
-
model_output,
|
| 558 |
-
sample,
|
| 559 |
-
timestep,
|
| 560 |
-
clip_denoised=clip_denoised,
|
| 561 |
-
generator=generator,
|
| 562 |
-
cond_grad=cond_grad,
|
| 563 |
-
)
|
| 564 |
-
|
| 565 |
-
prev_sample = out["sample"]
|
| 566 |
-
pred_original_sample = out.get("pred_xstart")
|
| 567 |
-
|
| 568 |
-
if not return_dict:
|
| 569 |
-
return (prev_sample, pred_original_sample)
|
| 570 |
-
|
| 571 |
-
return ADMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
|
| 572 |
-
|
| 573 |
-
def create_runtime(self, num_inference_steps: Optional[int] = None, use_ddim: bool = False) -> _SpacedDiffusion:
|
| 574 |
-
"""
|
| 575 |
-
Build a spaced diffusion object for legacy loop-based sampling (`p_sample_loop`).
|
| 576 |
-
|
| 577 |
-
Prefer `set_timesteps` + `step` for Diffusers-style inference.
|
| 578 |
-
"""
|
| 579 |
-
timestep_respacing = self.config.timestep_respacing
|
| 580 |
-
if num_inference_steps is not None:
|
| 581 |
-
timestep_respacing = f"ddim{num_inference_steps}" if use_ddim else str(num_inference_steps)
|
| 582 |
-
return _create_spaced_diffusion(
|
| 583 |
-
steps=self.config.steps,
|
| 584 |
-
learn_sigma=self.config.learn_sigma,
|
| 585 |
-
sigma_small=self.config.sigma_small,
|
| 586 |
-
noise_schedule=self.config.noise_schedule,
|
| 587 |
-
predict_xstart=self.config.predict_xstart,
|
| 588 |
-
rescale_timesteps=self.config.rescale_timesteps,
|
| 589 |
-
timestep_respacing=timestep_respacing,
|
| 590 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|