File size: 13,423 Bytes
be89dda |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from .layers import ResidualBlock, AttnBlock
from .utils import get_named_beta_schedule
def sinusoidal_embedding(n, d):
"""
n: iteration steps,
d: time embedding dimension
"""
# Returns the standard positional embedding
embedding = torch.tensor([[i / 10000 ** (2 * j / d) for j in range(d)] for i in range(n)])
sin_mask = torch.arange(0, n, 2)
embedding[sin_mask] = torch.sin(embedding[sin_mask])
embedding[1 - sin_mask] = torch.cos(embedding[sin_mask])
return embedding
def _make_te(dim_in, dim_out):
return nn.Sequential(
nn.Linear(dim_in, dim_out),
nn.SiLU(),
nn.Linear(dim_out, dim_out)
)
class UNet_with_time(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
input_frame = config.input_frame
output_frame = config.output_frame
n_steps = config.n_steps
time_emb_dim = config.time_emb_dim
cond_nc = config.cond_nc
chs_mult = config.chs_mult ## e.g. (1, 2, 4, 8)
n_res_blocks = config.n_res_blocks
base_chs = config.base_chs
## e.g. (0, 0, 1, 1) -> 0 means no attention
use_attn_list = config.use_attn_list
layer_depth = len(chs_mult)
assert len(use_attn_list) == layer_depth, "length of use_attn_list should be the same as chs_mult"
assert input_frame >= output_frame, "input_frame should be larger than or equal to output_frame"
self.filter_list = [base_chs * m for m in chs_mult]
## time embedding
self.time_embed = nn.Embedding(n_steps, time_emb_dim)
self.time_embed.weight.data = sinusoidal_embedding(n_steps, time_emb_dim)
self.time_embed.requires_grad_(False)
self.time_embed_fc = _make_te(time_emb_dim, time_emb_dim)
## end of time embedding
## input conv
self.input_layer = nn.PixelUnshuffle(downscale_factor=2)
## downsampling
self.down_blocks = nn.ModuleList()
in_c = input_frame * 4 ## after pixel unshuffle
for i in range(layer_depth):
out_c = self.filter_list[i]
for _ in range(n_res_blocks):
self.down_blocks.append(
ResidualBlock(in_c, in_c, cond_nc, time_emb_dim, down_flag=False, up_flag=False)
)
if use_attn_list[i]:
self.down_blocks.append(AttnBlock(in_c, 4)) ## num_head=4
self.down_blocks.append(
ResidualBlock(in_c, out_c, cond_nc, time_emb_dim, down_flag=True, up_flag=False)
)
in_c = out_c
## end of downsampling
## middle
self.mid_block1 = ResidualBlock(in_c, in_c, cond_nc, time_emb_dim, down_flag=False, up_flag=False)
self.mid_attn = AttnBlock(in_c, 4)
self.mid_block2 = ResidualBlock(in_c, in_c, cond_nc, time_emb_dim, down_flag=False, up_flag=False)
## end of middle
## upsampling
self.up_blocks = nn.ModuleList()
self.filter_list = [input_frame * 4] + self.filter_list[:-1]
for i in reversed(range(layer_depth)): ## i = layer_depth-1, ..., 0
out_c = self.filter_list[i]
self.up_blocks.append(
ResidualBlock(in_c*2, out_c, cond_nc, time_emb_dim, down_flag=False, up_flag=True)
)
if use_attn_list[i]:
self.up_blocks.append(AttnBlock(out_c)) ## num_head=1
for _ in range(n_res_blocks):
self.up_blocks.append(
ResidualBlock(out_c*2, out_c, cond_nc, time_emb_dim, down_flag=False, up_flag=False)
)
in_c = out_c
## end of upsampling
self.out_up = nn.PixelShuffle(upscale_factor=2)
self.out_conv = nn.Conv2d(input_frame, output_frame, 3, padding=1)
def forward(self, x, t, cond):
"""
x: (b, in_c, h, w), noisy input (concatenated with some data)
t: (b,), time step
cond: (b, cond_nc, h, w), conditional input
"""
# time embedding
t_emb = self.time_embed(t) ## (b, time_emb_dim)
t_emb = self.time_embed_fc(t_emb) ## (b, time_emb_dim)
# input conv
x = self.input_layer(x)
# downsampling
skip_x = []
for ii, down_layer in enumerate(self.down_blocks):
if isinstance(down_layer, ResidualBlock):
x = down_layer(x, cond, t_emb)
skip_x.append(x)
elif isinstance(down_layer, AttnBlock):
x = down_layer(x)
else:
raise ValueError("Wrong layer type in down_blocks")
# middle
x = self.mid_block1(x, cond, t_emb)
x = self.mid_attn(x)
x = self.mid_block2(x, cond, t_emb)
# upsampling
for up_layer in self.up_blocks:
if isinstance(up_layer, ResidualBlock):
skip_feat = skip_x.pop()
x = torch.cat([x, skip_feat], dim=1) ## concat along channel dimension
x = up_layer(x, cond, t_emb)
elif isinstance(up_layer, AttnBlock):
x = up_layer(x)
else:
raise ValueError("Wrong layer type in up_blocks")
# output
x = self.out_up(x)
x = self.out_conv(x)
return x
class DDPM(nn.Module):
def __init__(self, backbone, output_shape, n_steps=1000, min_beta=1e-4, max_beta=0.02, device='cuda'):
"""
output_shape: dim(C, H, W)
"""
super().__init__()
self.device = device
self.backbone_model = backbone
self.output_shape = output_shape
self.n_steps = n_steps
## linear betas
betas = get_named_beta_schedule("linear", n_steps, min_beta, max_beta)
alphas = 1.0 - betas
alpha_bars = torch.cumprod(alphas, dim=0)
self.register_buffer('betas', betas)
self.register_buffer('alphas', alphas)
self.register_buffer('alpha_bars', alpha_bars)
def forward(self, x, t, cond):
"""
x: (b, in_c, h, w), noisy input (concatenated with some data)
cond: (b, cond_nc, h, w), conditional input
t: (b,), time step
"""
return self.backbone_model(x, t, cond)
@torch.no_grad()
def add_noise(self, x0, t, eta=None):
"""
x0: (b, c, h, w), original data
t: (b,), time step (0 <= t < n_steps)
"""
b, c, h, w = x0.shape
if eta is None:
eta = torch.randn(b, c, h, w, device=x0.device)
alpha_bar = self.alpha_bars[t]
noisy_x = alpha_bar.sqrt().reshape(b, 1, 1, 1) * x0 + (1 - alpha_bar).sqrt().reshape(b, 1, 1, 1) * eta
return noisy_x
def denoise(self, xt, t, cond):
"""
xt: (b, in_c, h, w), noisy input (concatenated with some data)
cond: (b, cond_nc, h, w), conditional input
t: (b,), time step (0 <= t < n_steps)
"""
pred_noise = self(xt, t, cond)
return pred_noise
@torch.no_grad()
def _build_progress_iter(self, iterable, total, mode: str):
"""
Internal helper to create a progress iterator based on verbose mode.
"""
mode = (mode or "none").lower()
if mode == "tqdm":
try:
from tqdm import tqdm
return tqdm(iterable, total=total, desc="DDPM sampling", leave=False), mode
except Exception:
return iterable, "none"
return iterable, mode
@torch.no_grad()
def sample_ddpm(self, cond, input_cond=None, verbose: str = "none", store_intermediate: bool = False):
"""
input_frame: (b, c, h, w) number of input frames (conditional input frames) for the diffusion model
cond: (b, cond_nc, h, w), conditional input
verbose: "none", "text", or "tqdm" for progress display
"""
## confirm that the model is in eval mode
self.backbone_model.eval()
B, C, H, W = cond.shape
## get cond device
device = cond.device
x = torch.randn(B, *self.output_shape, device=device)
progress_iter_raw = reversed(range(self.n_steps))
progress_iter, mode = self._build_progress_iter(progress_iter_raw, self.n_steps, verbose)
use_text = mode == "text"
text_interval = max(1, self.n_steps // 10)
frames = []
for idx, t in enumerate(progress_iter):
time_tensor = (torch.ones(B, device=device) * t).long()
if input_cond is not None:
input_ = torch.cat((x, input_cond), dim=1)
else:
input_ = x
eta_theta = self.denoise(input_, time_tensor, cond)
alpha_t = self.alphas[t]
alpha_t_bar = self.alpha_bars[t]
a = 1 / alpha_t.sqrt()
b = ((1 - alpha_t) / (1 - alpha_t_bar).sqrt()) * eta_theta
x = a * (x - b)
if t > 0:
z = torch.randn(B, *self.output_shape, device=device)
beta_t = self.betas[t]
sigma_t = beta_t.sqrt()
x = x + sigma_t * z
## store intermediate frames for visualization
if (idx % 50 == 0) or (t == 0):
out = x.clone()
out = ((out + 1) / 2).clamp(0, 1)
out = out.cpu().numpy()
frames.append(out)
if use_text and (idx + 1) % text_interval == 0:
print(f"DDPM sampling {idx + 1}/{self.n_steps}", flush=True)
if mode == "tqdm" and hasattr(progress_iter, "close"):
progress_iter.close()
if store_intermediate:
return x, frames
else:
return x
@torch.no_grad()
def sample_ddim(self, cond, input_cond=None, ddim_steps: int = 100, eta: float = 0.2, verbose: str = "none", store_intermediate: bool = False):
"""
Deterministic/stochastic DDIM sampling.
cond: (b, cond_nc, h, w)
input_cond: optional conditional input concatenated with the predicted frames
ddim_steps: number of steps to sample (<= n_steps)
eta: 0 for deterministic DDIM, >0 adds noise controlled by eta
verbose: "none", "text", or "tqdm" for progress display
"""
self.backbone_model.eval()
B, C, H, W = cond.shape
device = cond.device
ddim_steps = max(1, min(ddim_steps, self.n_steps))
# create evenly spaced timesteps
ddim_timesteps = torch.linspace(0, self.n_steps - 1, steps=ddim_steps, device=device).long()
ddim_timesteps = torch.unique(ddim_timesteps, sorted=True) # safety against duplicates
ddim_t_reverse = list(reversed(ddim_timesteps.tolist()))
x = torch.randn(B, *self.output_shape, device=device)
progress_iter_raw = enumerate(ddim_t_reverse)
progress_iter, mode = self._build_progress_iter(progress_iter_raw, len(ddim_t_reverse), verbose)
use_text = mode == "text"
text_interval = max(1, len(ddim_t_reverse) // 10)
frames = []
for idx, (iter_idx, t) in enumerate(progress_iter):
time_tensor = torch.full((B,), t, device=device, dtype=torch.long)
if input_cond is not None:
input_ = torch.cat((x, input_cond), dim=1)
else:
input_ = x
eps = self.denoise(input_, time_tensor, cond)
alpha_bar_t = self.alpha_bars[t]
sqrt_alpha_bar_t = alpha_bar_t.sqrt()
sqrt_one_minus_alpha_bar_t = (1 - alpha_bar_t).sqrt()
x0_pred = (x - sqrt_one_minus_alpha_bar_t * eps) / sqrt_alpha_bar_t
if iter_idx + 1 < len(ddim_t_reverse):
t_prev = ddim_t_reverse[iter_idx + 1]
alpha_bar_prev = self.alpha_bars[t_prev]
else:
alpha_bar_prev = torch.ones_like(alpha_bar_t, device=device)
sigma_t = 0.0
if eta > 0 and alpha_bar_prev < 1:
sigma_t = eta * torch.sqrt(
(1 - alpha_bar_prev) / (1 - alpha_bar_t) * (1 - alpha_bar_t / alpha_bar_prev)
)
sigma_t = torch.as_tensor(sigma_t, device=device, dtype=x.dtype)
noise = torch.randn_like(x) if (eta > 0 and alpha_bar_prev < 1) else torch.zeros_like(x)
c_t = torch.sqrt(torch.clamp(1 - alpha_bar_prev - sigma_t ** 2, min=0.0))
x = (
alpha_bar_prev.sqrt() * x0_pred
+ c_t * eps
+ sigma_t * noise
)
## store intermediate frames for visualization
if (idx % 25 == 0) or (t == 0):
out = x.clone()
out = ((out + 1) / 2).clamp(0, 1)
out = out.cpu().numpy()
frames.append(out)
if use_text and (idx + 1) % text_interval == 0:
print(f"DDIM sampling {idx + 1}/{len(ddim_t_reverse)}", flush=True)
if mode == "tqdm" and hasattr(progress_iter, "close"):
progress_iter.close()
if store_intermediate:
return x, frames
else:
return x
# Backward-compatible alias
sample = sample_ddpm
|