File size: 8,148 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 | """
Denoising Diffusion Probabilistic Models — Evaluation (FID Score)
Paper: https://arxiv.org/abs/2006.11239
Authors: Ho, Jain, Abbeel (2020)
§4 — "We report FID score and Inception Score... Our best results
are FID: 3.17, IS: 9.46 (unconditional) on CIFAR-10."
This module provides a wrapper for sampling from a trained DDPM model
and computing FID scores using the pytorch-fid library.
FID (Fréchet Inception Distance) is the primary evaluation metric
used in §4 and Table 1. Lower FID = better quality.
NOTE: FID computation requires:
1. Generating 50K samples (standard for CIFAR-10)
2. Computing Inception features for real and generated images
3. Computing the Fréchet distance between the two feature distributions
This is computationally expensive. For quick validation, generate a
small batch and visually inspect.
"""
import os
import logging
from pathlib import Path
from typing import Optional
import torch
import torchvision
import yaml
from model import UNet, UNetConfig
from utils import linear_noise_schedule, sample, EMA
logging.basicConfig(level=logging.INFO, format="%(asctime)s — %(message)s")
logger = logging.getLogger(__name__)
def load_model(
checkpoint_path: str,
device: torch.device,
use_ema: bool = True,
) -> tuple:
"""Load a trained DDPM model from checkpoint.
§4 — "we also report results with an exponential moving average"
The EMA parameters typically produce better samples.
Args:
checkpoint_path: Path to .pt checkpoint file
device: Target device
use_ema: Whether to load EMA parameters (recommended)
Returns:
(model, config_dict)
"""
checkpoint = torch.load(checkpoint_path, map_location=device, weights_only=False)
cfg = checkpoint["config"]
model_cfg = cfg["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=cfg["data"].get("image_size", 32),
)
model = UNet(unet_config).to(device)
if use_ema and "ema_state_dict" in checkpoint:
# Load EMA parameters
ema_params = checkpoint["ema_state_dict"]
for name, param in model.named_parameters():
if name in ema_params:
param.data.copy_(ema_params[name])
logger.info("Loaded EMA parameters")
else:
model.load_state_dict(checkpoint["model_state_dict"])
logger.info("Loaded model parameters (no EMA)")
return model, cfg
@torch.no_grad()
def generate_samples(
model: UNet,
config: dict,
num_samples: int = 64,
device: Optional[torch.device] = None,
) -> torch.Tensor:
"""Generate samples using Algorithm 2 — Sampling.
§3.4 — "sampling from p_θ(x_{t-1} | x_t) = N(x_{t-1}; μ_θ(x_t, t), σ²_t I)"
Args:
model: Trained UNet model
config: Config dict from checkpoint
num_samples: Number of images to generate
device: Target device
Returns:
(num_samples, C, H, W) — generated images in [0, 1] range
"""
if device is None:
device = next(model.parameters()).device
diff_cfg = config["diffusion"]
data_cfg = config["data"]
T = diff_cfg["T"]
betas = linear_noise_schedule(T, diff_cfg["beta_start"], diff_cfg["beta_end"]).to(device)
image_size = data_cfg.get("image_size", 32)
image_channels = config["model"].get("image_channels", 3)
shape = (num_samples, image_channels, image_size, image_size)
model.eval()
samples = sample(model, shape, T, betas, device)
# Convert from [-1, 1] to [0, 1] for saving
samples = (samples + 1.0) / 2.0
samples = samples.clamp(0.0, 1.0)
return samples
def save_samples(
samples: torch.Tensor,
output_dir: str,
prefix: str = "sample",
make_grid: bool = True,
nrow: int = 8,
):
"""Save generated samples as images.
Args:
samples: (N, C, H, W) in [0, 1]
output_dir: Directory to save images
prefix: Filename prefix
make_grid: If True, also save a grid image
nrow: Number of images per row in grid
"""
output_dir = Path(output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
if make_grid:
grid = torchvision.utils.make_grid(samples, nrow=nrow, padding=2)
grid_path = output_dir / f"{prefix}_grid.png"
torchvision.utils.save_image(grid, grid_path)
logger.info(f"Saved grid: {grid_path}")
# Save individual images
for i, img in enumerate(samples):
img_path = output_dir / f"{prefix}_{i:05d}.png"
torchvision.utils.save_image(img, img_path)
logger.info(f"Saved {len(samples)} individual images to {output_dir}")
def compute_fid(
generated_dir: str,
real_stats_path: Optional[str] = None,
batch_size: int = 50,
device: str = "cuda",
dims: int = 2048,
) -> float:
"""Compute FID score between generated samples and real data.
§4 — "We report FID score... Our best results are FID: 3.17"
Requires the pytorch-fid package: pip install pytorch-fid
For CIFAR-10, you need pre-computed stats for the real training set,
or provide a directory of real images.
Args:
generated_dir: Directory containing generated .png images
real_stats_path: Path to pre-computed .npz stats for real data,
OR directory containing real images
batch_size: Batch size for Inception feature extraction
device: Device for computation
dims: Inception feature dimensionality (2048 = pool3)
Returns:
FID score (float). Lower is better.
"""
try:
from pytorch_fid import fid_score
except ImportError:
logger.error(
"pytorch-fid not installed. Install with: pip install pytorch-fid\n"
"Then re-run evaluation."
)
raise
if real_stats_path is None:
raise ValueError(
"Must provide real_stats_path: either a .npz file with pre-computed "
"Inception statistics, or a directory of real CIFAR-10 images."
)
fid = fid_score.calculate_fid_given_paths(
[generated_dir, real_stats_path],
batch_size=batch_size,
device=torch.device(device),
dims=dims,
)
logger.info(f"FID score: {fid:.2f}")
return fid
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser(description="DDPM Evaluation — Generate samples and compute FID")
parser.add_argument("--checkpoint", type=str, required=True,
help="Path to trained model checkpoint")
parser.add_argument("--num_samples", type=int, default=64,
help="Number of samples to generate")
parser.add_argument("--output_dir", type=str, default="generated",
help="Output directory for generated images")
parser.add_argument("--fid", action="store_true",
help="Compute FID score (requires --real_stats)")
parser.add_argument("--real_stats", type=str, default=None,
help="Path to real data stats (.npz) or directory")
parser.add_argument("--no_ema", action="store_true",
help="Don't use EMA parameters")
parser.add_argument("--device", type=str, default="cuda",
help="Device (cuda/cpu)")
args = parser.parse_args()
device = torch.device(args.device if torch.cuda.is_available() else "cpu")
model, cfg = load_model(args.checkpoint, device, use_ema=not args.no_ema)
samples = generate_samples(model, cfg, args.num_samples, device)
save_samples(samples, args.output_dir)
if args.fid:
compute_fid(args.output_dir, args.real_stats, device=args.device)
|