File size: 9,159 Bytes
6835659 | 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 | """
Improved image generator with domain gating, similarity thresholding,
and explicit retrieval failure reporting.
Phase 1B+1C: Addresses retrieval reliability for controlled experiments.
"""
from __future__ import annotations
import logging
from dataclasses import dataclass
from pathlib import Path
from typing import List, Optional, Tuple
import numpy as np
from src.embeddings.aligned_embeddings import AlignedEmbedder
from src.embeddings.similarity import cosine_similarity
from src.exceptions import IndexError_
logger = logging.getLogger(__name__)
# Domain keywords for gating — reject obvious mismatches
DOMAIN_KEYWORDS = {
"nature": {"forest", "tree", "mountain", "jungle", "garden", "park", "field",
"meadow", "countryside", "rural", "fog", "dawn", "sunrise", "hill",
"valley", "woodland", "grove", "leaf", "green", "wildlife"},
"urban": {"city", "street", "neon", "urban", "downtown", "skyscraper",
"building", "traffic", "night", "cobblestone", "road", "car",
"sign", "shop", "window", "concrete", "sidewalk"},
"water": {"beach", "ocean", "wave", "sea", "shore", "coast", "lake",
"river", "water", "sand", "surf", "tide", "tropical", "island"},
}
# Domains that should NOT co-occur in prompt+image
INCOMPATIBLE_DOMAINS = {
"nature": {"urban"},
"urban": {"nature", "water"},
"water": {"urban"},
}
@dataclass
class ImageRetrievalResult:
"""Result of image retrieval with metadata for experiment bundles."""
image_path: str
similarity: float
domain: str
retrieval_failed: bool
candidates_considered: int
candidates_above_threshold: int
top_5: List[Tuple[str, float]]
def _detect_prompt_domain(prompt: str) -> Optional[str]:
"""Detect the primary domain of a prompt from keywords."""
prompt_lower = prompt.lower()
prompt_words = set(prompt_lower.split())
scores = {}
for domain, keywords in DOMAIN_KEYWORDS.items():
overlap = len(prompt_words & keywords)
# Also check substring matches for compound words
substring_hits = sum(1 for kw in keywords if kw in prompt_lower)
scores[domain] = overlap + substring_hits
if not scores or max(scores.values()) == 0:
return None
best_domain = max(scores, key=scores.get)
return best_domain
def _is_domain_compatible(prompt_domain: Optional[str], image_domain: str) -> bool:
"""Check if image domain is compatible with prompt domain."""
if prompt_domain is None:
return True # No domain detected, allow everything
if image_domain == "other":
return True # Unknown domain, don't reject
incompatible = INCOMPATIBLE_DOMAINS.get(prompt_domain, set())
return image_domain not in incompatible
class ImprovedImageRetrievalGenerator:
"""
Image retrieval with:
- Domain gating: rejects obvious domain mismatches (forest prompt → no city images)
- Raised similarity floor: min_similarity=0.20 (was 0.15)
- Explicit retrieval failure: returns retrieval_failed=True instead of silent nonsense
- Full diagnostic metadata for experiment bundles
"""
def __init__(
self,
index_path: str = "data/embeddings/image_index.npz",
min_similarity: float = 0.20,
top_k: int = 5,
):
self.index_path = Path(index_path)
self.min_similarity = min_similarity
self.top_k = top_k
if not self.index_path.exists():
raise IndexError_(
f"Missing image index at {self.index_path}. "
"Run: python scripts/build_embedding_indexes.py",
index_path=str(self.index_path),
)
data = np.load(self.index_path, allow_pickle=True)
self.ids = data["ids"].tolist()
self.embs = data["embs"].astype("float32")
# Load domain tags if available (from rebuilt index)
if "domains" in data:
self.domains = data["domains"].tolist()
else:
# Infer from filenames for old indexes
self.domains = [self._infer_domain(p) for p in self.ids]
if len(self.ids) == 0:
raise IndexError_(
"Image index is empty. "
"Add images and run: python scripts/build_embedding_indexes.py",
index_path=str(self.index_path),
)
self.embedder = AlignedEmbedder(target_dim=512)
@staticmethod
def _infer_domain(filepath: str) -> str:
"""Infer domain from filename."""
name = Path(filepath).stem.lower()
for domain, keywords in DOMAIN_KEYWORDS.items():
if any(kw in name for kw in keywords):
return domain
return "other"
def retrieve(
self,
query_text: str,
min_similarity: Optional[float] = None,
) -> ImageRetrievalResult:
"""
Retrieve best matching image with domain gating and quality checks.
Returns ImageRetrievalResult with full metadata including retrieval_failed flag.
"""
if min_similarity is None:
min_similarity = self.min_similarity
prompt_domain = _detect_prompt_domain(query_text)
query_emb = self.embedder.embed_text(query_text)
# Score all candidates
scored = []
for img_path, img_emb, img_domain in zip(self.ids, self.embs, self.domains):
sim = cosine_similarity(query_emb, img_emb)
scored.append((img_path, sim, img_domain))
scored.sort(key=lambda x: x[1], reverse=True)
top_5 = [(Path(p).name, s) for p, s, _ in scored[:5]]
# Phase 1: Domain gating — filter out incompatible domains
domain_filtered = [
(p, s, d) for p, s, d in scored
if _is_domain_compatible(prompt_domain, d)
]
# Phase 2: Similarity thresholding
candidates = domain_filtered if domain_filtered else scored
above_threshold = [(p, s, d) for p, s, d in candidates if s >= min_similarity]
if above_threshold:
# Best candidate passes both domain and similarity checks
best_path, best_sim, best_domain = above_threshold[0]
return ImageRetrievalResult(
image_path=best_path,
similarity=best_sim,
domain=best_domain,
retrieval_failed=False,
candidates_considered=len(scored),
candidates_above_threshold=len(above_threshold),
top_5=top_5,
)
# Phase 3: Fallback — nothing passed threshold
# Return best domain-compatible candidate (even if below threshold)
if domain_filtered:
best_path, best_sim, best_domain = domain_filtered[0]
else:
best_path, best_sim, best_domain = scored[0]
return ImageRetrievalResult(
image_path=best_path,
similarity=best_sim,
domain=best_domain,
retrieval_failed=best_sim < min_similarity,
candidates_considered=len(scored),
candidates_above_threshold=0,
top_5=top_5,
)
# Backward-compatible method
def retrieve_top_k(
self,
query_text: str,
k: int = 1,
min_similarity: Optional[float] = None,
) -> List[Tuple[str, float]]:
"""Backward-compatible interface. Returns list of (path, score) tuples."""
result = self.retrieve(query_text, min_similarity)
return [(result.image_path, result.similarity)]
def generate_image_improved(
prompt: str,
out_dir: str,
index_path: str = "data/embeddings/image_index.npz",
min_similarity: float = 0.20,
) -> str:
"""
Generate (retrieve) an image for a prompt.
Returns the image path. Warns on low similarity or retrieval failure.
"""
generator = ImprovedImageRetrievalGenerator(
index_path=index_path,
min_similarity=min_similarity,
)
result = generator.retrieve(prompt, min_similarity=min_similarity)
if result.retrieval_failed:
logger.warning(
"Image retrieval failed: no image above threshold (%.2f) "
"for prompt: \"%s...\" — best: %s (sim=%.4f, domain=%s)",
min_similarity, prompt[:60], Path(result.image_path).name,
result.similarity, result.domain,
)
elif result.similarity < 0.25:
logger.warning(
"Low image similarity: %.4f for \"%s...\" -> %s",
result.similarity, prompt[:60], Path(result.image_path).name,
)
return result.image_path
def generate_image_with_metadata(
prompt: str,
index_path: str = "data/embeddings/image_index.npz",
min_similarity: float = 0.20,
) -> ImageRetrievalResult:
"""
Generate (retrieve) an image and return full metadata.
Use this in experiment pipelines where retrieval quality matters.
"""
generator = ImprovedImageRetrievalGenerator(
index_path=index_path,
min_similarity=min_similarity,
)
return generator.retrieve(prompt, min_similarity=min_similarity)
|