File size: 11,876 Bytes
c61c435 | 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 | from __future__ import annotations
from dataclasses import asdict, dataclass, field
from datetime import datetime, timezone
import hashlib
import json
from pathlib import Path
from typing import Any, Callable, Sequence
from adam.eve import EveVisionModel, classify_eve_embeddings
RATINGS = {"favorite", "keep", "unsure", "reject"}
POSITIVE_RATINGS = {"favorite", "keep"}
NEGATIVE_RATINGS = {"reject"}
def _now() -> str:
return datetime.now(timezone.utc).isoformat()
def preference_profile_id(provider_id: str, model_path: str) -> str:
key = f"{provider_id}\n{str(Path(model_path).expanduser().resolve())}"
return hashlib.sha1(key.encode("utf-8")).hexdigest()[:16]
def image_cache_id(path: str | Path) -> str:
resolved = str(Path(path).expanduser().resolve())
return hashlib.sha1(resolved.encode("utf-8")).hexdigest()
@dataclass(slots=True)
class GenerationRating:
image_path: str
rating: str
provider_id: str
model_name: str
model_path: str
seed: int = 0
sampler: str = ""
steps: int = 0
resolution: str = ""
generation_settings: dict[str, Any] = field(default_factory=dict)
generation_created_at: str = ""
rated_at: str = field(default_factory=_now)
embedding: list[float] | None = None
@classmethod
def from_dict(cls, payload: dict[str, Any]) -> "GenerationRating":
rating = str(payload.get("rating", "unsure")).casefold()
return cls(
image_path=str(Path(str(payload.get("image_path", ""))).expanduser().resolve()),
rating=rating if rating in RATINGS else "unsure",
provider_id=str(payload.get("provider_id", "")),
model_name=str(payload.get("model_name", "")),
model_path=str(payload.get("model_path", "")),
seed=int(payload.get("seed", 0) or 0),
sampler=str(payload.get("sampler", "")),
steps=int(payload.get("steps", 0) or 0),
resolution=str(payload.get("resolution", "")),
generation_settings=dict(payload.get("generation_settings") or {}),
generation_created_at=str(payload.get("generation_created_at", "")),
rated_at=str(payload.get("rated_at") or _now()),
embedding=[float(value) for value in payload["embedding"]]
if isinstance(payload.get("embedding"), list)
else None,
)
@dataclass(frozen=True, slots=True)
class PreferenceScore:
image_path: str
score: float | None
confidence: float
category: str
reason: str = ""
class PreferenceProfile:
def __init__(self, root: Path, provider_id: str, model_name: str, model_path: str) -> None:
self.root = root.resolve()
self.provider_id = provider_id
self.model_name = model_name
self.model_path = str(Path(model_path).expanduser().resolve()) if model_path else ""
self.id = preference_profile_id(provider_id, self.model_path)
self.path = self.root / "data" / "generation_preferences" / f"{self.id}.json"
self.keep_threshold = 0.70
self.reject_threshold = 0.35
self.ratings: dict[str, GenerationRating] = {}
self.load()
def load(self) -> None:
try:
payload = json.loads(self.path.read_text(encoding="utf-8"))
except (OSError, ValueError, TypeError, json.JSONDecodeError):
return
self.model_name = str(payload.get("model_name") or self.model_name)
self.provider_id = str(payload.get("provider_id") or self.provider_id)
self.model_path = str(payload.get("model_path") or self.model_path)
thresholds = payload.get("thresholds", {})
if isinstance(thresholds, dict):
self.keep_threshold = float(thresholds.get("keep", self.keep_threshold))
self.reject_threshold = float(thresholds.get("reject", self.reject_threshold))
ratings = payload.get("ratings", [])
if isinstance(ratings, list):
for item in ratings:
if isinstance(item, dict):
rating = GenerationRating.from_dict(item)
self.ratings[rating.image_path] = rating
def save(self) -> None:
self.path.parent.mkdir(parents=True, exist_ok=True)
temporary = self.path.with_suffix(".tmp")
temporary.write_text(
json.dumps(
{
"version": 1,
"profile_id": self.id,
"provider_id": self.provider_id,
"model_name": self.model_name,
"model_path": self.model_path,
"thresholds": {
"keep": self.keep_threshold,
"reject": self.reject_threshold,
},
"ratings": [asdict(item) for item in self.ratings.values()],
"updated_at": _now(),
},
indent=2,
),
encoding="utf-8",
)
temporary.replace(self.path)
def set_rating(
self,
image_path: str | Path,
rating: str,
*,
seed: int = 0,
sampler: str = "",
steps: int = 0,
resolution: str = "",
generation_settings: dict[str, Any] | None = None,
generation_created_at: str = "",
embedding: Sequence[float] | None = None,
) -> GenerationRating:
clean = rating.casefold().strip()
if clean not in RATINGS:
raise ValueError("Generation rating must be Favorite, Keep, Unsure, or Reject.")
resolved = str(Path(image_path).expanduser().resolve())
existing = self.ratings.get(resolved)
record = GenerationRating(
image_path=resolved,
rating=clean,
provider_id=self.provider_id,
model_name=self.model_name,
model_path=self.model_path,
seed=int(seed),
sampler=sampler,
steps=int(steps),
resolution=resolution,
generation_settings=dict(generation_settings or {}),
generation_created_at=generation_created_at,
rated_at=_now(),
embedding=[float(value) for value in embedding] if embedding is not None else (
existing.embedding if existing else None
),
)
self.ratings[resolved] = record
self.save()
return record
def rating_for(self, image_path: str | Path) -> GenerationRating | None:
return self.ratings.get(str(Path(image_path).expanduser().resolve()))
def examples(self) -> tuple[list[GenerationRating], list[GenerationRating]]:
positive = [
item for item in self.ratings.values()
if item.rating in POSITIVE_RATINGS and Path(item.image_path).is_file()
]
negative = [
item for item in self.ratings.values()
if item.rating in NEGATIVE_RATINGS and Path(item.image_path).is_file()
]
return positive, negative
def has_signal(self) -> bool:
positive, _negative = self.examples()
return bool(positive)
class ImageEmbeddingCache:
def __init__(self, root: Path, model_id: str) -> None:
self.root = root.resolve()
self.model_id = model_id
self.folder = self.root / "data" / "image_embeddings" / hashlib.sha1(model_id.encode("utf-8")).hexdigest()[:12]
def get(self, path: str | Path) -> list[float] | None:
cache_path = self.folder / f"{image_cache_id(path)}.json"
try:
payload = json.loads(cache_path.read_text(encoding="utf-8"))
except (OSError, ValueError, TypeError, json.JSONDecodeError):
return None
source = Path(path).expanduser().resolve()
try:
stat = source.stat()
except OSError:
return None
if payload.get("path") != str(source) or payload.get("mtime") != stat.st_mtime:
return None
vector = payload.get("embedding")
return [float(value) for value in vector] if isinstance(vector, list) else None
def set(self, path: str | Path, embedding: Sequence[float]) -> None:
source = Path(path).expanduser().resolve()
try:
stat = source.stat()
except OSError:
return
self.folder.mkdir(parents=True, exist_ok=True)
cache_path = self.folder / f"{image_cache_id(source)}.json"
temporary = cache_path.with_suffix(".tmp")
temporary.write_text(
json.dumps(
{
"path": str(source),
"mtime": stat.st_mtime,
"model_id": self.model_id,
"embedding": [float(value) for value in embedding],
}
),
encoding="utf-8",
)
temporary.replace(cache_path)
class GenerationPreferenceEvaluator:
"""Shared EVE-backed scorer for generated images."""
def __init__(
self,
root: Path,
vision: EveVisionModel | None = None,
*,
embedder: Callable[[Sequence[str | Path]], list[list[float]]] | None = None,
) -> None:
self.root = root.resolve()
self.vision = vision or EveVisionModel(prefer_gpu=False)
self.embedder = embedder
self.cache = ImageEmbeddingCache(self.root, self.vision.model_id)
def _embedding(self, path: str | Path) -> list[float]:
cached = self.cache.get(path)
if cached is not None:
return cached
vectors = self.embedder([path]) if self.embedder else self.vision.embed([path])
vector = [float(value) for value in vectors[0]]
self.cache.set(path, vector)
return vector
def score(
self,
profile: PreferenceProfile,
image_paths: Sequence[str | Path],
*,
keep_threshold: float | None = None,
reject_threshold: float | None = None,
) -> list[PreferenceScore]:
positive, negative = profile.examples()
if not positive:
return [
PreferenceScore(str(Path(path).expanduser().resolve()), None, 0.0, "Needs Review", "No preference examples yet")
for path in image_paths
]
positive_vectors = [item.embedding or self._embedding(item.image_path) for item in positive]
negative_vectors = [item.embedding or self._embedding(item.image_path) for item in negative]
image_vectors = [self._embedding(path) for path in image_paths]
keep = max(0.001, min(1.0, float(keep_threshold if keep_threshold is not None else profile.keep_threshold)))
reject = max(0.0, min(float(reject_threshold if reject_threshold is not None else profile.reject_threshold), keep - 0.001))
results = classify_eve_embeddings(
image_paths,
image_vectors,
positive_vectors,
negative_vectors,
keep_threshold=keep,
reject_threshold=reject,
)
categories = {"keep": "Strong Keep", "reject": "Likely Reject", "unreviewed": "Needs Review"}
return [
PreferenceScore(result.path, result.match_score, result.decision_confidence, categories[result.suggestion])
for result in results
]
def score_generated_images(
root: Path,
*,
provider_id: str,
model_name: str,
model_path: str,
image_paths: Sequence[str | Path],
keep_threshold: float | None = None,
reject_threshold: float | None = None,
) -> list[PreferenceScore]:
profile = PreferenceProfile(root, provider_id, model_name, model_path)
evaluator = GenerationPreferenceEvaluator(root)
return evaluator.score(
profile,
image_paths,
keep_threshold=keep_threshold,
reject_threshold=reject_threshold,
)
|