File size: 6,634 Bytes
256c9c2 | 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 | """
Denoising Diffusion Probabilistic Models — Training Loop
Paper: https://arxiv.org/abs/2006.11239
Authors: Ho, Jain, Abbeel (2020)
Implements: Algorithm 1 — Training procedure
1: repeat
2: x_0 ~ q(x_0) ← sample data
3: t ~ Uniform({1, ..., T}) ← sample timestep
4: ε ~ N(0, I) ← sample noise
5: Take gradient step on
∇_θ ||ε − ε_θ(√ᾱ_t x_0 + √(1−ᾱ_t) ε, t)||²
6: until converged
Hyperparameters from §4 and Appendix B:
- Adam optimizer, learning rate 2e-4 (§B)
- No lr schedule/warmup [UNSPECIFIED — not mentioned in paper]
- Gradient clipping at 1.0 [FROM_OFFICIAL_CODE]
- EMA with decay 0.9999 (§4)
- T = 1000 (§4)
- Trained for 800K steps on CIFAR-10 (Appendix B)
"""
import os
import logging
from pathlib import Path
import torch
import torch.nn as nn
import yaml
from model import UNet, UNetConfig
from loss import DDPMLoss
from data import get_dataloaders
from utils import linear_noise_schedule, q_sample, EMA
logging.basicConfig(level=logging.INFO, format="%(asctime)s — %(message)s")
logger = logging.getLogger(__name__)
def train(config_path: str = "configs/base.yaml"):
"""Algorithm 1 — DDPM Training.
Args:
config_path: Path to YAML config file
"""
# --- Load config ---
config_path = Path(config_path)
if config_path.exists():
with open(config_path) as f:
cfg = yaml.safe_load(f)
else:
raise FileNotFoundError(f"Config not found: {config_path}")
diff_cfg = cfg["diffusion"]
model_cfg = cfg["model"]
train_cfg = cfg["training"]
data_cfg = cfg["data"]
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
logger.info(f"Using device: {device}")
# --- Noise schedule ---
# §2, Eq. 4 — linear schedule β_1 = 0.0001, β_T = 0.02
T = diff_cfg["T"]
betas = linear_noise_schedule(T, diff_cfg["beta_start"], diff_cfg["beta_end"])
betas = betas.to(device)
alphas = 1.0 - betas # α_t = 1 − β_t
alpha_bar = torch.cumprod(alphas, dim=0) # ᾱ_t = ∏_{s=1}^{t} α_s
sqrt_alpha_bar = torch.sqrt(alpha_bar) # √ᾱ_t
sqrt_one_minus_alpha_bar = torch.sqrt(1.0 - alpha_bar) # √(1−ᾱ_t)
# --- Model ---
unet_config = UNetConfig(
image_channels=model_cfg.get("image_channels", 3),
base_channels=model_cfg.get("base_channels", 128),
channel_mults=tuple(model_cfg.get("channel_mults", [1, 2, 2, 2])),
num_res_blocks=model_cfg.get("num_res_blocks", 2),
attention_resolutions=tuple(model_cfg.get("attention_resolutions", [16])),
dropout=model_cfg.get("dropout", 0.0),
num_groups=model_cfg.get("num_groups", 32),
image_size=data_cfg.get("image_size", 32),
)
model = UNet(unet_config).to(device)
logger.info(f"Model: {model}")
# --- EMA ---
# §4 — "we also report results with an exponential moving average of
# model parameters with a decay factor of 0.9999"
ema = EMA(model, decay=train_cfg.get("ema_decay", 0.9999))
# --- Optimizer ---
# Appendix B — "Adam, lr = 2 × 10^-4"
optimizer = torch.optim.Adam(
model.parameters(),
lr=float(train_cfg.get("lr", 2e-4)),
)
# --- Loss ---
criterion = DDPMLoss()
# --- Data ---
train_loader, _ = get_dataloaders(
data_dir=data_cfg.get("data_dir", "./data"),
batch_size=train_cfg.get("batch_size", 128),
num_workers=data_cfg.get("num_workers", 4),
image_size=data_cfg.get("image_size", 32),
)
# --- Training loop: Algorithm 1 ---
total_steps = train_cfg.get("total_steps", 800_000)
log_every = train_cfg.get("log_every", 1000)
save_every = train_cfg.get("save_every", 50_000)
save_dir = Path(train_cfg.get("save_dir", "checkpoints"))
save_dir.mkdir(parents=True, exist_ok=True)
grad_clip = train_cfg.get("gradient_clip", 1.0)
step = 0
model.train()
while step < total_steps:
for batch in train_loader:
if step >= total_steps:
break
# Algorithm 1, line 2: x_0 ~ q(x_0)
x_0 = batch[0].to(device) # (batch, C, H, W), labels discarded (unconditional)
batch_size = x_0.shape[0]
# Algorithm 1, line 3: t ~ Uniform({1, ..., T})
t = torch.randint(1, T + 1, (batch_size,), device=device)
# Algorithm 1, line 4: ε ~ N(0, I)
noise = torch.randn_like(x_0)
# Algorithm 1, line 5: compute x_t and predict noise
x_t = q_sample(x_0, t, sqrt_alpha_bar, sqrt_one_minus_alpha_bar, noise)
noise_pred = model(x_t, t)
# L_simple — §3.4, Eq. 14
loss = criterion(noise_pred, noise)
# Algorithm 1, line 5: gradient step
optimizer.zero_grad()
loss.backward()
# [FROM_OFFICIAL_CODE] gradient clipping
if grad_clip > 0:
nn.utils.clip_grad_norm_(model.parameters(), grad_clip)
optimizer.step()
# §4 — EMA update
ema.update()
step += 1
if step % log_every == 0:
logger.info(f"Step {step}/{total_steps} — loss: {loss.item():.6f}")
if step % save_every == 0:
checkpoint = {
"step": step,
"model_state_dict": model.state_dict(),
"ema_state_dict": ema.shadow_params,
"optimizer_state_dict": optimizer.state_dict(),
"loss": loss.item(),
"config": cfg,
}
ckpt_path = save_dir / f"ddpm_step_{step}.pt"
torch.save(checkpoint, ckpt_path)
logger.info(f"Saved checkpoint: {ckpt_path}")
# Final save
final_path = save_dir / "ddpm_final.pt"
torch.save({
"step": step,
"model_state_dict": model.state_dict(),
"ema_state_dict": ema.shadow_params,
"optimizer_state_dict": optimizer.state_dict(),
"config": cfg,
}, final_path)
logger.info(f"Training complete. Final checkpoint: {final_path}")
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser(description="Train DDPM — Algorithm 1")
parser.add_argument("--config", type=str, default="configs/base.yaml",
help="Path to config YAML")
args = parser.parse_args()
train(args.config)
|