Datasets:
File size: 9,441 Bytes
b0d8bdd | 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 | # Copyright 2026 Nikhil Raghavendra
# SPDX-License-Identifier: Apache-2.0
"""Reference evaluation for SEN-Bridge.
Usage::
python scripts/evaluate.py SAMPLES.npy --data DATASET_DIR_OR_REPO
[--split test] [--fid-stats stats/fid_test.npz] [--framework torch|jax]
[--bootstrap 1000] [--out results.json]
``SAMPLES.npy`` holds one generated optical image per example of the split, in
dataset order: uint8 ``[N, 128, 128, 3]`` or float in [-1, 1]. The script reports
FID (against the statistics file, or against the split's targets if none is given)
and per-image PSNR, SSIM and LPIPS as mean and std, with 95% bootstrap intervals
resampled by geographic cluster. Never use the test split to choose models.
"""
from __future__ import annotations
import argparse
import json
from pathlib import Path
from typing import TYPE_CHECKING, Any, Protocol
import numpy as np
from datasets import load_dataset
if TYPE_CHECKING:
from numpy.typing import NDArray
_HALF_RANGE = 127.5
_METRICS = ("psnr", "ssim", "lpips")
_CI = (2.5, 97.5)
class Backend(Protocol):
"""A metrics backend (PyTorch or JAX) behind a common interface."""
def paired(
self, gen: NDArray[Any], tgt: NDArray[Any]
) -> tuple[NDArray[Any], NDArray[Any], NDArray[Any]]:
"""Score generated images against targets.
Args:
gen: Generated images, uint8 or float in [-1, 1], ``[B, H, W, 3]``.
tgt: Target images, same shape.
Returns:
PSNR, SSIM and LPIPS for each image.
"""
...
def feed(self, images: NDArray[Any], *, real: bool) -> None:
"""Add images to the FID accumulator.
Args:
images: Images, uint8 or float in [-1, 1], ``[B, H, W, 3]``.
real: Whether they are real targets.
"""
...
def fid(self) -> float:
"""Compute the FID.
Returns:
The Frechet Inception Distance.
"""
...
def to_float(x: NDArray[Any]) -> NDArray[np.float32]:
"""Map images to float32 in [-1, 1].
Args:
x: uint8 images, or floats that are already in [-1, 1].
Returns:
Float32 images in [-1, 1].
"""
if x.dtype == np.uint8:
return x.astype(np.float32) / _HALF_RANGE - 1
return x.astype(np.float32)
class TorchBackend:
"""Metrics from ``sen_bridge.torch``."""
def __init__(self, stats: Path | None) -> None:
"""Create the backend.
Args:
stats: Real-target statistics file, or ``None`` to featurise the
targets.
"""
# torch is an optional dependency, imported only when selected.
import torch # ruff: ignore[import-outside-top-level]
# Same optional dependency as the import above.
from sen_bridge import torch as sb # ruff: ignore[import-outside-top-level]
self._torch, self._sb = torch, sb
self._fid = sb.FID.from_stats(stats) if stats else sb.FID()
# Returns a framework tensor, which has no common static type.
def _prep(self, x: NDArray[Any]) -> Any: # ruff: ignore[any-type]
batch = self._torch.from_numpy(to_float(x)).permute(0, 3, 1, 2)
return batch.to(
self._fid.device
) # LPIPS-VGG on 30k pairs is far too slow on CPU
def paired(
self, gen: NDArray[Any], tgt: NDArray[Any]
) -> tuple[NDArray[Any], NDArray[Any], NDArray[Any]]:
"""Score generated images against targets.
Args:
gen: Generated images, ``[B, H, W, 3]``.
tgt: Target images, same shape.
Returns:
PSNR, SSIM and LPIPS for each image.
"""
a, b = self._prep(gen), self._prep(tgt)
return (
self._sb.psnr(a, b).cpu().numpy(),
self._sb.ssim(a, b).cpu().numpy(),
self._sb.lpips(a, b).cpu().numpy(),
)
def feed(self, images: NDArray[Any], *, real: bool) -> None:
"""Add images to the FID accumulator.
Args:
images: Images, ``[B, H, W, 3]``.
real: Whether they are real targets.
"""
self._fid.update(self._prep(images), real=real)
def fid(self) -> float:
"""Compute the FID.
Returns:
The Frechet Inception Distance.
"""
return self._fid.compute()
class JaxBackend:
"""Metrics from ``sen_bridge.jax``."""
def __init__(self, stats: Path | None) -> None:
"""Create the backend.
Args:
stats: Real-target statistics file, or ``None`` to featurise the
targets.
"""
# jax is an optional dependency, imported only when selected.
import jax # ruff: ignore[import-outside-top-level]
# Same optional dependency as the import above.
import jax.numpy as jnp # ruff: ignore[import-outside-top-level]
# Same optional dependency as the import above.
from sen_bridge import jax as sb # ruff: ignore[import-outside-top-level]
self._jnp = jnp
self._fns = [jax.jit(jax.vmap(f)) for f in (sb.psnr, sb.ssim, sb.lpips)]
self._fid = sb.FID.from_stats(stats) if stats else sb.FID()
def paired(
self, gen: NDArray[Any], tgt: NDArray[Any]
) -> tuple[NDArray[Any], NDArray[Any], NDArray[Any]]:
"""Score generated images against targets.
Args:
gen: Generated images, ``[B, H, W, 3]``.
tgt: Target images, same shape.
Returns:
PSNR, SSIM and LPIPS for each image.
"""
a, b = self._jnp.asarray(to_float(gen)), self._jnp.asarray(to_float(tgt))
psnr, ssim, lpips = (np.asarray(f(a, b)) for f in self._fns)
return psnr, ssim, lpips
def feed(self, images: NDArray[Any], *, real: bool) -> None:
"""Add images to the FID accumulator.
Args:
images: Images, ``[B, H, W, 3]``.
real: Whether they are real targets.
"""
self._fid.update(self._jnp.asarray(to_float(images)), real=real)
def fid(self) -> float:
"""Compute the FID.
Returns:
The Frechet Inception Distance.
"""
return self._fid.compute()
def cluster_bootstrap(
values: dict[str, NDArray[Any]],
clusters: NDArray[Any],
n: int,
) -> dict[str, tuple[float, float]]:
"""Bootstrap 95% intervals of per-image means, resampling whole clusters.
Args:
values: Per-image scores by metric name.
clusters: Cluster id of each image.
n: Number of bootstrap resamples.
Returns:
The lower and upper bound for each metric.
"""
unique = np.unique(clusters)
members = {c: np.flatnonzero(clusters == c) for c in unique}
rng = np.random.default_rng(0)
draws = []
for _ in range(n):
picked = np.concatenate([
members[c] for c in rng.choice(unique, len(unique))
])
draws.append([v[picked].mean() for v in values.values()])
low, high = np.percentile(draws, _CI, axis=0)
return {
k: (float(a), float(b)) for k, a, b in zip(values, low, high, strict=True)
}
def main() -> None:
"""Evaluate a set of generated samples and print the metrics as JSON.
Raises:
SystemExit: If the number of samples differs from the split size.
"""
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("samples", type=Path)
parser.add_argument("--data", required=True)
parser.add_argument("--split", default="test")
parser.add_argument("--fid-stats", type=Path)
parser.add_argument("--framework", choices=["torch", "jax"], default="torch")
parser.add_argument("--bootstrap", type=int, default=1000)
parser.add_argument("--batch", type=int, default=100)
parser.add_argument("--out", type=Path)
args = parser.parse_args()
ds = load_dataset(args.data, split=args.split)
generated = np.load(args.samples)
if len(generated) != len(ds):
msg = f"{len(generated)} samples for {len(ds)} examples"
raise SystemExit(msg)
backend: Backend = (
TorchBackend(args.fid_stats)
if args.framework == "torch"
else JaxBackend(args.fid_stats)
)
scores: dict[str, list[NDArray[Any]]] = {m: [] for m in _METRICS}
for start in range(0, len(ds), args.batch):
stop = min(start + args.batch, len(ds))
targets = np.stack([
np.asarray(r["optical"]) for r in ds.select(range(start, stop))
])
gen = generated[start:stop]
for name, value in zip(_METRICS, backend.paired(gen, targets), strict=True):
scores[name].append(value)
backend.feed(gen, real=False)
if args.fid_stats is None:
backend.feed(targets, real=True)
per_image = {k: np.concatenate(v) for k, v in scores.items()}
intervals = cluster_bootstrap(
per_image, np.asarray(ds["cluster_id"]), args.bootstrap
)
result: dict[str, Any] = {"n": len(ds), "fid": backend.fid()}
for name, values in per_image.items():
result[name] = {
"mean": float(values.mean()),
"std": float(values.std()),
"ci95_cluster_bootstrap": list(intervals[name]),
}
print(json.dumps(result, indent=2))
if args.out:
args.out.write_text(json.dumps(result, indent=2), encoding="utf-8")
if __name__ == "__main__":
main()
|