NinadJoshi382 commited on
Commit
390bcbd
·
verified ·
1 Parent(s): dbdd188

Upload ET_all_well.py

Browse files
Files changed (1) hide show
  1. ET_all_well.py +1429 -0
ET_all_well.py ADDED
@@ -0,0 +1,1429 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ EntryTrial.py
3
+
4
+ A fully self-contained, single-file reimplementation of ArtArena's
5
+ Entry-Trials pipeline:
6
+
7
+ prep_ET.py -> ET_infer.py -> ET_eval_fixed.py
8
+ S
9
+ No dependency on the ArtArena repo directory — everything the three
10
+ original scripts did is reproduced in-process below. Supports sweeping
11
+ over multiple datasets, multiple T2I models, and multiple proximity
12
+ metrics in one call.
13
+
14
+ Built-in models: sd15, sdxl, sana15 (as in the original ET_infer.py).
15
+ A clearly marked PLACEHOLDER section lets you register any additional
16
+ text-to-image model — see "ADD NEW MODELS HERE" below.
17
+
18
+ CSD metric: expects a local ./CSD folder next to this script, containing
19
+ model.py and utils.py (exactly as you've set it up). Override with
20
+ --csd_repo_dir if you keep it elsewhere.
21
+
22
+ -------------------------------------------------------------------------
23
+ Examples
24
+ -------------------------------------------------------------------------
25
+
26
+ Single dataset, single model, single metric:
27
+
28
+ python EntryTrial.py \
29
+ --datasets /content/dataset \
30
+ --models sd15 \
31
+ --metrics clip \
32
+ --output_root /content/saves \
33
+ --num_imgs 2
34
+
35
+ Sweep: two datasets x three models x two metrics, resumable:
36
+
37
+ python EntryTrial.py \
38
+ --datasets /content/dataset_a /content/dataset_b \
39
+ --models sd15 sdxl sana15 \
40
+ --metrics clip lpips \
41
+ --output_root /content/saves \
42
+ --num_imgs 2 \
43
+ --skip_existing
44
+
45
+ CSD metric (reads ./CSD/model.py + ./CSD/utils.py next to this script):
46
+
47
+ python EntryTrial.py \
48
+ --datasets /content/dataset \
49
+ --models sd15 \
50
+ --metrics csd \
51
+ --csd_model_path /content/CSD/checkpoint.pt \
52
+ --output_root /content/saves \
53
+ --num_imgs 2
54
+
55
+ List the models currently registered:
56
+
57
+ python EntryTrial.py --list_models
58
+ """
59
+
60
+ from __future__ import annotations
61
+
62
+ import argparse
63
+ import contextlib
64
+ import csv
65
+ import gc
66
+ import hashlib
67
+ import importlib
68
+ import json
69
+ import re
70
+ import shutil
71
+ import sys
72
+ import time
73
+ from dataclasses import dataclass, field
74
+ from pathlib import Path
75
+ from typing import Any, Callable, Dict, List, Optional, Tuple
76
+
77
+ import torch
78
+ from PIL import Image, ImageFile
79
+
80
+ import contextlib
81
+
82
+ @contextlib.contextmanager
83
+ def temporary_sys_path(path: str):
84
+ """Temporarily prepend `path` to sys.path, then restore it exactly as it was."""
85
+ sys.path.insert(0, path)
86
+ try:
87
+ yield
88
+ finally:
89
+ sys.path.remove(path)
90
+
91
+ ImageFile.LOAD_TRUNCATED_IMAGES = True # don't error on truncated images
92
+
93
+ SCRIPT_DIR = Path(__file__).resolve().parent
94
+ IMAGE_EXTENSIONS = {".jpg", ".jpeg", ".png", ".bmp", ".tiff", ".tif", ".webp"}
95
+ VALID_IMAGE_EXTS = (".jpg", ".jpeg", ".png", ".tiff", ".webp")
96
+
97
+
98
+ # =============================================================================
99
+ # STAGE 1 — dataset scanning / prompt + mapping construction (was prep_ET.py)
100
+ # =============================================================================
101
+
102
+ @dataclass
103
+ class PromptRecord:
104
+ orig_path: str
105
+ artist: str
106
+ artwork: str
107
+ prompt: str
108
+ gen_paths: List[str]
109
+
110
+
111
+ def parse_artwork_artist(stem: str, split_on: str = "last") -> Tuple[str, str]:
112
+ """
113
+ LEGACY positional parser, kept only as the final fallback for files the
114
+ frequency-based method below can't confidently resolve (e.g. a genuinely
115
+ rare artist with too few paintings to clear --min_artist_count). Prefer
116
+ build_artist_artwork_mapping() for the actual dataset scan.
117
+ """
118
+ if "_" not in stem:
119
+ return stem.replace("_", " "), "unknown"
120
+
121
+ if split_on == "first":
122
+ idx = stem.index("_")
123
+ else: # "last"
124
+ idx = stem.rindex("_")
125
+ artwork_raw, artist_raw = stem[:idx], stem[idx + 1:]
126
+
127
+ return artwork_raw.replace("_", " ").strip(), artist_raw.replace("_", " ").strip()
128
+
129
+
130
+ # =============================================================================
131
+ # Frequency-based artist/artwork parsing
132
+ # =============================================================================
133
+ # A real artist's name recurs as the EXACT SAME trailing underscore-token
134
+ # sequence across every one of their paintings (dozens to hundreds of files),
135
+ # while a one-off quirk in a title (a Bible-verse reference like
136
+ # "Genesis 22_2", a colon-substitute like "Composition_ The Guitar", a
137
+ # duplicate-title disambiguation suffix) is, by construction, unique to a
138
+ # single file. That frequency gap is detectable directly from the dataset's
139
+ # filenames, with no external artist list and no access to the original
140
+ # per-artist source folder structure required.
141
+
142
+ def _clean_artwork_text(s: str) -> str:
143
+ return re.sub(r"\s+", " ", s).strip()
144
+
145
+
146
+ def build_suffix_counts(stems: List[str], max_tokens: int) -> Dict[Tuple[str, ...], int]:
147
+ counts: Dict[Tuple[str, ...], int] = {}
148
+ for stem in stems:
149
+ toks = stem.split("_")
150
+ for n in range(1, min(max_tokens, len(toks) - 1) + 1):
151
+ key = tuple(toks[-n:])
152
+ counts[key] = counts.get(key, 0) + 1
153
+ return counts
154
+
155
+
156
+ def _find_artist_token_count(stem: str, counts: Dict[Tuple[str, ...], int], max_tokens: int, min_count: int) -> Optional[int]:
157
+ toks = stem.split("_")
158
+ best_n = None
159
+ for n in range(1, min(max_tokens, len(toks) - 1) + 1):
160
+ if counts.get(tuple(toks[-n:]), 0) >= min_count:
161
+ best_n = n # counts are non-increasing in n; keep the largest that still clears
162
+ return best_n
163
+
164
+
165
+ def build_artist_artwork_mapping(
166
+ stems: List[str], max_tokens: int = 4, min_count: int = 3
167
+ ) -> Tuple[Dict[str, Dict[str, str]], List[str]]:
168
+ """
169
+ Returns ({stem: {"artwork": ..., "artist": ...}}, unresolved_stems).
170
+ Two passes: raw stems first, then a retry with a trailing "_<digits>"
171
+ collision-suffix stripped for anything unresolved on the first pass
172
+ (recovers cases like "..._Claude_Monet_2" without disturbing genuine
173
+ title-content digits, which normally appear BEFORE the artist and are
174
+ unaffected by stripping only the trailing end of the string).
175
+ """
176
+ counts = build_suffix_counts(stems, max_tokens)
177
+ results: Dict[str, Dict[str, str]] = {}
178
+ unresolved: List[str] = []
179
+
180
+ for stem in stems:
181
+ n = _find_artist_token_count(stem, counts, max_tokens, min_count)
182
+ if n is None:
183
+ unresolved.append(stem)
184
+ continue
185
+ toks = stem.split("_")
186
+ artist = " ".join(toks[-n:])
187
+ artwork = _clean_artwork_text(" ".join(toks[:-n]))
188
+ results[stem] = {"artwork": artwork, "artist": artist}
189
+
190
+ if not unresolved:
191
+ return results, []
192
+
193
+ def strip_trailing_digits(s: str) -> str:
194
+ return re.sub(r"_\d+$", "", s)
195
+
196
+ stripped_stems = [strip_trailing_digits(s) for s in stems]
197
+ stripped_counts = build_suffix_counts(stripped_stems, max_tokens)
198
+
199
+ still_unresolved: List[str] = []
200
+ for stem in unresolved:
201
+ stripped = strip_trailing_digits(stem)
202
+ if stripped == stem:
203
+ still_unresolved.append(stem)
204
+ continue
205
+ n = _find_artist_token_count(stripped, stripped_counts, max_tokens, min_count)
206
+ if n is None:
207
+ still_unresolved.append(stem)
208
+ continue
209
+ toks = stripped.split("_")
210
+ artist = " ".join(toks[-n:])
211
+ artwork = _clean_artwork_text(" ".join(toks[:-n]))
212
+ results[stem] = {"artwork": artwork, "artist": artist}
213
+
214
+ return results, still_unresolved
215
+
216
+
217
+ def save_artist_artwork_mapping(mapping: Dict[str, Dict[str, str]], path: Path) -> None:
218
+ path.parent.mkdir(parents=True, exist_ok=True)
219
+ with open(path, "w", encoding="utf-8") as f:
220
+ json.dump(mapping, f, indent=2, ensure_ascii=False)
221
+ print(f"[OK] Saved artist/artwork mapping ({len(mapping)} entries) -> {path}")
222
+
223
+
224
+ def load_artist_artwork_mapping(path: Path) -> Dict[str, Dict[str, str]]:
225
+ with open(path, "r", encoding="utf-8") as f:
226
+ return json.load(f)
227
+
228
+
229
+ def collect_images(dataset_dir: Path) -> List[Path]:
230
+ """All image files directly inside dataset_dir (non-recursive, sorted)."""
231
+ return sorted(
232
+ p for p in dataset_dir.iterdir()
233
+ if p.is_file() and p.suffix.lower() in IMAGE_EXTENSIONS
234
+ )
235
+
236
+
237
+ def build_prompt(artwork: str, artist: str) -> str:
238
+ return f"{artwork} in the style of {artist}"
239
+
240
+
241
+ def sanitize_name(name: str) -> str:
242
+ """Make a string safe for file/folder names (spaces -> underscores, strip unsafe chars)."""
243
+ name = name.strip()
244
+ name = re.sub(r"\s", "_", name)
245
+ name = re.sub(r"[^A-Za-z0-9_.-]", "", name)
246
+ return name
247
+
248
+
249
+ def generated_paths_for(artist: str, artwork: str, root_output_dir: Path, num_imgs: int) -> List[str]:
250
+ """
251
+ root_output_dir/<artist>_<artwork>_<idx>.png, 1-based, zero-padded to 2 digits.
252
+
253
+ NOTE: sanitize_name() is applied here — the SAME single place used both to
254
+ build the mapping.json entries and to save the actual generated files
255
+ later, so the two can never drift apart (the original two-script version
256
+ sanitized only in ET_infer.py, not in prep_ET.py's mapping — a latent
257
+ mismatch bug if artist/artwork ever contained characters like ' or : ).
258
+ """
259
+ artist_slug = sanitize_name(artist)
260
+ artwork_slug = sanitize_name(artwork)
261
+ return [
262
+ str(root_output_dir / f"{artist_slug}_{artwork_slug}_{i + 1:02d}.png")
263
+ for i in range(num_imgs)
264
+ ]
265
+
266
+
267
+ def build_prompts_and_mapping(
268
+ dataset_dir: Path, root_output_dir: Path, num_imgs: int, name_split: str,
269
+ artist_artwork_mapping_path: Optional[Path] = None,
270
+ min_artist_count: int = 3, max_artist_tokens: int = 4,
271
+ ) -> List[PromptRecord]:
272
+ """
273
+ name_split is kept for backward compatibility (used only as the final
274
+ fallback for anything the frequency method can't resolve) but is no
275
+ longer the primary parsing method.
276
+
277
+ If artist_artwork_mapping_path already exists on disk, it's loaded and
278
+ reused as-is (so hand-corrections persist and don't get silently
279
+ overwritten by a recompute). Otherwise it's built fresh from THIS
280
+ dataset_dir's full file list — the frequency method needs the full
281
+ dataset in view to detect which trailing token sequences recur often
282
+ enough to be a real artist, so this mapping should be built once at the
283
+ dataset level and reused across every model that dataset is run with
284
+ (not recomputed per model).
285
+ """
286
+ if not dataset_dir.exists() or not dataset_dir.is_dir():
287
+ sys.exit(f"[ERROR] dataset_dir does not exist or is not a directory: {dataset_dir}")
288
+
289
+ images = collect_images(dataset_dir)
290
+ if not images:
291
+ sys.exit(f"[ERROR] No supported image files found in: {dataset_dir}")
292
+
293
+ stems = [img.stem for img in images]
294
+
295
+ if artist_artwork_mapping_path is not None and artist_artwork_mapping_path.is_file():
296
+ artist_artwork_map = load_artist_artwork_mapping(artist_artwork_mapping_path)
297
+ print(f"[i] Reusing existing artist/artwork mapping -> {artist_artwork_mapping_path} "
298
+ f"({len(artist_artwork_map)} entries). Delete it to force a fresh frequency-based rebuild.")
299
+ unresolved_stems = [s for s in stems if s not in artist_artwork_map]
300
+ else:
301
+ artist_artwork_map, unresolved_stems = build_artist_artwork_mapping(
302
+ stems, max_tokens=max_artist_tokens, min_count=min_artist_count
303
+ )
304
+ artist_counts: Dict[str, int] = {}
305
+ for entry in artist_artwork_map.values():
306
+ artist_counts[entry["artist"]] = artist_counts.get(entry["artist"], 0) + 1
307
+ print(f"[i] Frequency-based parsing discovered {len(artist_counts)} distinct artist(s):")
308
+ for artist, count in sorted(artist_counts.items(), key=lambda kv: -kv[1]):
309
+ print(f" {artist!r:30s} {count} file(s)")
310
+ if artist_artwork_mapping_path is not None:
311
+ save_artist_artwork_mapping(artist_artwork_map, artist_artwork_mapping_path)
312
+
313
+ if unresolved_stems:
314
+ print(f"[!] {len(unresolved_stems)} file(s) didn't clear --min_artist_count via the frequency "
315
+ f"method (rare artist, or a naming quirk) — falling back to positional split for these:")
316
+ for stem in unresolved_stems[:20]:
317
+ print(f" {stem}")
318
+ if len(unresolved_stems) > 20:
319
+ print(f" ... and {len(unresolved_stems) - 20} more")
320
+
321
+ records: List[PromptRecord] = []
322
+ skipped = 0
323
+ for img_path in images:
324
+ entry = artist_artwork_map.get(img_path.stem)
325
+ if entry is not None:
326
+ artwork, artist = entry["artwork"], entry["artist"]
327
+ else:
328
+ artwork, artist = parse_artwork_artist(img_path.stem, split_on=name_split)
329
+
330
+ if not artwork or not artist or artist == "unknown":
331
+ print(f" [!] Skipping '{img_path.name}': could not parse artwork/artist.")
332
+ skipped += 1
333
+ continue
334
+ prompt = build_prompt(artwork, artist)
335
+ gen_paths = generated_paths_for(artist, artwork, root_output_dir, num_imgs)
336
+ records.append(PromptRecord(str(img_path), artist, artwork, prompt, gen_paths))
337
+
338
+ if not records:
339
+ sys.exit("[ERROR] No valid image/prompt pairs could be built.")
340
+ if skipped:
341
+ print(f"[!] Skipped {skipped} file(s) due to parse failures.")
342
+ print(f"[i] Built {len(records)} prompt record(s) from {dataset_dir}")
343
+ return records
344
+
345
+
346
+ def save_prompts_pt(records: List[PromptRecord], pt_path: Path) -> None:
347
+ pt_path.parent.mkdir(parents=True, exist_ok=True)
348
+ torch.save([r.prompt for r in records], str(pt_path))
349
+ print(f"[OK] Saved {len(records)} prompt(s) -> {pt_path}")
350
+
351
+
352
+ def save_mapping_json(records: List[PromptRecord], json_path: Path) -> None:
353
+ json_path.parent.mkdir(parents=True, exist_ok=True)
354
+ mapping = {r.orig_path: r.gen_paths for r in records}
355
+ with open(json_path, "w", encoding="utf-8") as f:
356
+ json.dump(mapping, f, indent=2, ensure_ascii=False)
357
+ print(f"[OK] Saved mapping ({len(mapping)} entries) -> {json_path}")
358
+
359
+
360
+ def load_existing_records(
361
+ pt_path: Path, json_path: Path, name_split: str,
362
+ artist_artwork_mapping: Optional[Dict[str, Dict[str, str]]] = None,
363
+ ) -> List[PromptRecord]:
364
+ """
365
+ Reconstructs the exact PromptRecord list from an already-saved
366
+ prompts.pt + mapping.json pair, WITHOUT rescanning the dataset or
367
+ touching either file. This is what lets multiple parallel
368
+ --start_index/--end_index jobs share one canonical prompt/mapping set:
369
+ only the job that actually built them (via 'prep' in --stages) writes
370
+ to disk; every other job just reads.
371
+
372
+ prompts.pt is a plain list[str] of prompts, saved in the same order as
373
+ mapping.json's keys (both come from the same `records` list in one
374
+ save_prompts_pt/save_mapping_json call), so zipping them back together
375
+ by position is safe as long as neither file was hand-edited to reorder
376
+ entries.
377
+
378
+ artist/artwork aren't stored in either file (only orig_path, prompt,
379
+ gen_paths are), but generate_images() needs them for its dedup key and
380
+ logging -- so they're re-derived via a plain O(1) dict lookup into the
381
+ SAME artist_artwork_mapping already used to build this prompt set (no
382
+ frequency recomputation), falling back to the positional parser only
383
+ for a stem that mapping doesn't cover.
384
+ """
385
+ if not pt_path.is_file() or not json_path.is_file():
386
+ raise FileNotFoundError(
387
+ f"'prep' excluded from --stages but prompts.pt/mapping.json not found "
388
+ f"({pt_path}, {json_path}). Run a job with 'prep' in --stages first."
389
+ )
390
+
391
+ prompts = torch.load(str(pt_path))
392
+ with open(json_path, "r", encoding="utf-8") as f:
393
+ mapping = json.load(f)
394
+
395
+ orig_paths = list(mapping.keys())
396
+ if len(orig_paths) != len(prompts):
397
+ raise ValueError(
398
+ f"prompts.pt ({len(prompts)} entries) and mapping.json ({len(orig_paths)} entries) "
399
+ f"are out of sync at {pt_path.parent} -- delete both and rerun with 'prep' in --stages."
400
+ )
401
+
402
+ records: List[PromptRecord] = []
403
+ for orig_path, prompt in zip(orig_paths, prompts):
404
+ stem = Path(orig_path).stem
405
+ entry = artist_artwork_mapping.get(stem) if artist_artwork_mapping else None
406
+ if entry is not None:
407
+ artwork, artist = entry["artwork"], entry["artist"]
408
+ else:
409
+ artwork, artist = parse_artwork_artist(stem, split_on=name_split)
410
+ records.append(PromptRecord(orig_path, artist, artwork, prompt, mapping[orig_path]))
411
+
412
+ print(f"[i] Loaded {len(records)} existing record(s) from {pt_path.name}/{json_path.name} (no rebuild).")
413
+ return records
414
+
415
+
416
+ # =============================================================================
417
+ # STAGE 2 — image generation (was ET_infer.py)
418
+ # =============================================================================
419
+
420
+ class DiffusersPipelineWrapper:
421
+ """Adapts any standard diffusers pipeline to the uniform .generate(...) interface."""
422
+
423
+ def __init__(self, pipe, extra_call_kwargs: dict = None):
424
+ self.pipe = pipe
425
+ self.extra_call_kwargs = extra_call_kwargs or {} # e.g. {"height": 1024, "width": 1024}
426
+
427
+ def generate(self, prompt, num_inference_steps, guidance_scale, num_images_per_prompt, generator):
428
+ result = self.pipe(
429
+ prompt=prompt,
430
+ num_inference_steps=num_inference_steps,
431
+ guidance_scale=guidance_scale,
432
+ num_images_per_prompt=num_images_per_prompt,
433
+ generator=generator,
434
+ **self.extra_call_kwargs,
435
+ )
436
+ return result.images # list[PIL.Image]
437
+
438
+
439
+ def load_sd15(device: torch.device):
440
+ from diffusers import StableDiffusionPipeline
441
+ # MODEL_ID = "stable-diffusion-v1-5/stable-diffusion-v1-5"
442
+ MODEL_ID = r"G:\T2I_models\SDv15"
443
+ pipe = StableDiffusionPipeline.from_pretrained(
444
+ MODEL_ID, torch_dtype=torch.float16, safety_checker=None
445
+ ).to(device)
446
+ print("[model] Stable Diffusion v1.5 loaded successfully!")
447
+ return DiffusersPipelineWrapper(pipe)
448
+
449
+
450
+ def load_sdxl(device: torch.device):
451
+ from diffusers import StableDiffusionXLPipeline
452
+ # MODEL_ID = "stabilityai/stable-diffusion-xl-base-1.0"
453
+ MODEL_ID = r"G:\T2I_models\SDXL"
454
+ pipe = StableDiffusionXLPipeline.from_pretrained(
455
+ MODEL_ID, torch_dtype=torch.float16
456
+ ).to(device)
457
+ print("[model] Stable Diffusion XL loaded successfully!")
458
+ return DiffusersPipelineWrapper(pipe)
459
+
460
+
461
+ def load_sana15(device: torch.device):
462
+ from diffusers import SanaPipeline
463
+ pipe = SanaPipeline.from_pretrained( # NOTE: fixed the original repo's "frm_pretrained" typo
464
+ "Efficient-Large-Model/SANA1.5_1.6B_1024px_diffusers", torch_dtype=torch.bfloat16
465
+ ).to(device)
466
+ pipe.text_encoder = pipe.text_encoder.to(torch.bfloat16)
467
+ pipe.vae = pipe.vae.to(torch.bfloat16)
468
+ print("[model] SANA 1.5 loaded successfully!")
469
+ return DiffusersPipelineWrapper(pipe)
470
+
471
+ def load_flux(device: torch.device):
472
+ from diffusers import DiffusionPipeline
473
+ # MODEL_ID = r"G:\T2I_models\FLUX2" # <-- change this to wherever you saved the Flux model
474
+ MODEL_ID = "diffusers/FLUX.2-dev-bnb-4bit"
475
+ pipe = DiffusionPipeline.from_pretrained(
476
+ MODEL_ID,
477
+ dtype=torch.bfloat16,
478
+ device_map="cuda",
479
+ )
480
+ print("[model] FLUX.2 loaded successfully!")
481
+ return DiffusersPipelineWrapper(pipe)
482
+
483
+ # def load_flux_dev(device: torch.device):
484
+ # from diffusers import FluxPipeline
485
+ # MODEL_ID = "black-forest-labs/FLUX.1-dev"
486
+ # pipe = FluxPipeline.from_pretrained(
487
+ # MODEL_ID, torch_dtype=torch.bfloat16
488
+ # ).to(device)
489
+ # print("[model] FLUX.1-dev loaded successfully!")
490
+ # return DiffusersPipelineWrapper(pipe)
491
+ def load_flux_dev(device: torch.device):
492
+ from diffusers import FluxPipeline
493
+ MODEL_ID = "black-forest-labs/FLUX.1-dev"
494
+ pipe = FluxPipeline.from_pretrained(
495
+ MODEL_ID,
496
+ dtype=torch.bfloat16, # updated from torch_dtype, see note below
497
+ token="...", # <-- your token as a string
498
+ ).to(device)
499
+ print("[model] FLUX.1-dev loaded successfully!")
500
+ return DiffusersPipelineWrapper(pipe)
501
+
502
+ def load_sd35_medium(device: torch.device):
503
+ from diffusers import StableDiffusion3Pipeline
504
+ MODEL_ID = "stabilityai/stable-diffusion-3.5-medium"
505
+ pipe = StableDiffusion3Pipeline.from_pretrained(
506
+ MODEL_ID,
507
+ dtype=torch.bfloat16,
508
+ token="...", # <-- only needed if you're not using huggingface-cli login
509
+ ).to(device)
510
+ print("[model] Stable Diffusion 3.5 Medium loaded successfully!")
511
+ return DiffusersPipelineWrapper(pipe)
512
+
513
+ def load_qwen_image(device: torch.device):
514
+ from diffusers import DiffusionPipeline
515
+ MODEL_ID = "Qwen/Qwen-Image"
516
+ pipe = DiffusionPipeline.from_pretrained(
517
+ MODEL_ID, torch_dtype=torch.bfloat16
518
+ ).to(device)
519
+ print("[model] Qwen-Image loaded successfully!")
520
+ return DiffusersPipelineWrapper(pipe)
521
+
522
+ def load_sd35_large(device: torch.device):
523
+ from diffusers import StableDiffusion3Pipeline
524
+ MODEL_ID = "stabilityai/stable-diffusion-3.5-large"
525
+ pipe = StableDiffusion3Pipeline.from_pretrained(
526
+ MODEL_ID,
527
+ dtype=torch.bfloat16,
528
+ token="...", # only needed if not using huggingface-cli login
529
+ ).to(device)
530
+ print("[model] Stable Diffusion 3.5 Large loaded successfully!")
531
+ return DiffusersPipelineWrapper(pipe)
532
+
533
+ def load_pixart_alpha(device: torch.device):
534
+ from diffusers import PixArtAlphaPipeline
535
+ MODEL_ID = "PixArt-alpha/PixArt-XL-2-1024-MS"
536
+ pipe = PixArtAlphaPipeline.from_pretrained(MODEL_ID, dtype=torch.bfloat16).to(device)
537
+ print("[model] PixArt-alpha loaded successfully!")
538
+ return DiffusersPipelineWrapper(pipe)
539
+
540
+ def load_pixart_sigma(device: torch.device):
541
+ from diffusers import PixArtAlphaPipeline
542
+ MODEL_ID = "PixArt-alpha/PixArt-Sigma-XL-2-1024-MS"
543
+ pipe = PixArtAlphaPipeline.from_pretrained(MODEL_ID, dtype=torch.bfloat16).to(device)
544
+ print("[model] PixArt-Sigma loaded successfully!")
545
+ return DiffusersPipelineWrapper(pipe)
546
+
547
+ def load_auraflow(device: torch.device):
548
+ from diffusers import AuraFlowPipeline
549
+ MODEL_ID = "fal/AuraFlow-v0.3"
550
+ pipe = AuraFlowPipeline.from_pretrained(MODEL_ID, dtype=torch.bfloat16).to(device)
551
+ print("[model] AuraFlow v0.3 loaded successfully!")
552
+ return DiffusersPipelineWrapper(pipe)
553
+
554
+ def load_kandinsky22(device: torch.device):
555
+ from diffusers import KandinskyV22CombinedPipeline
556
+ MODEL_ID = "kandinsky-community/kandinsky-2-2-decoder"
557
+ pipe = KandinskyV22CombinedPipeline.from_pretrained(MODEL_ID, dtype=torch.bfloat16).to(device)
558
+ print("[model] Kandinsky 2.2 loaded successfully!")
559
+ return DiffusersPipelineWrapper(pipe)
560
+
561
+ class Kandinsky31Wrapper:
562
+ def __init__(self, t2i_pipe):
563
+ self.t2i_pipe = t2i_pipe
564
+
565
+ def generate(self, prompt, num_inference_steps, guidance_scale, num_images_per_prompt, generator):
566
+ return self.t2i_pipe(
567
+ prompt,
568
+ images_num=num_images_per_prompt,
569
+ steps=num_inference_steps,
570
+ guidance_scale=guidance_scale,
571
+ )
572
+
573
+
574
+ def load_kandinsky31(device: torch.device):
575
+ KANDINSKY3_REPO_PATH = r"G:\Kandinsky-3-main"
576
+
577
+ with temporary_sys_path(KANDINSKY3_REPO_PATH):
578
+ from kandinsky3 import get_T2I_pipeline
579
+
580
+ device_map = device
581
+ dtype_map = {
582
+ 'unet': torch.float16,
583
+ 'text_encoder': torch.float16,
584
+ 'movq': torch.float16,
585
+ }
586
+ t2i_pipe = get_T2I_pipeline(device_map, dtype_map,cache_dir= r"D:\kandinsky3")
587
+ print("[model] Kandinsky 3.1 (Full) loaded successfully!")
588
+ return Kandinsky31Wrapper(t2i_pipe)
589
+
590
+ class Kandinsky5Wrapper:
591
+ """
592
+ Kandinsky5T2IPipeline returns `.image` (singular) instead of `.images`
593
+ like most diffusers pipelines. Everything else matches the standard
594
+ .generate(...) interface DiffusersPipelineWrapper uses.
595
+ """
596
+ def __init__(self, pipe):
597
+ self.pipe = pipe
598
+
599
+ def generate(self, prompt, num_inference_steps, guidance_scale, num_images_per_prompt, generator):
600
+ result = self.pipe(
601
+ prompt=prompt,
602
+ num_inference_steps=num_inference_steps,
603
+ guidance_scale=guidance_scale,
604
+ num_images_per_prompt=num_images_per_prompt,
605
+ generator=generator,
606
+ )
607
+ return result.image # note: .image, not .images
608
+
609
+
610
+ def load_kandinsky5(device: torch.device):
611
+ from diffusers import Kandinsky5T2IPipeline
612
+ MODEL_ID = "kandinskylab/Kandinsky-5.0-T2I-Lite-sft-Diffusers"
613
+ pipe = Kandinsky5T2IPipeline.from_pretrained(
614
+ MODEL_ID, dtype=torch.bfloat16
615
+ ).to(device)
616
+ print("[model] Kandinsky 5.0 loaded successfully!")
617
+ return Kandinsky5Wrapper(pipe)
618
+
619
+ def load_z_image(device: torch.device):
620
+ from diffusers import ZImagePipeline
621
+ MODEL_ID = "Tongyi-MAI/Z-Image"
622
+ pipe = ZImagePipeline.from_pretrained(MODEL_ID, torch_dtype=torch.float16, low_cpu_mem_usage=False).to(device)
623
+ print("[model] Z-Image (standard/base) loaded successfully!")
624
+ return DiffusersPipelineWrapper(pipe, extra_call_kwargs={"height": 1024, "width": 1024})
625
+
626
+ class HunyuanImageWrapper:
627
+ """
628
+ HunyuanImagePipeline doesn't accept `guidance_scale` as a __call__ kwarg —
629
+ it's configured via a separate `guider` object instead (pipe.guider.new(...)).
630
+ It's also natively a 2K model, so height/width are fixed here rather than
631
+ left to the pipeline's own defaults. Everything else matches the same
632
+ .generate(...) interface DiffusersPipelineWrapper uses.
633
+ """
634
+ def __init__(self, pipe, height: int = 2048, width: int = 2048):
635
+ self.pipe = pipe
636
+ self.height = height
637
+ self.width = width
638
+
639
+ def generate(self, prompt, num_inference_steps, guidance_scale, num_images_per_prompt, generator):
640
+ # Route guidance_scale into the guider instead of passing it to __call__.
641
+ self.pipe.guider = self.pipe.guider.new(guidance_scale=guidance_scale)
642
+ result = self.pipe(
643
+ prompt=prompt,
644
+ num_inference_steps=num_inference_steps,
645
+ num_images_per_prompt=num_images_per_prompt,
646
+ generator=generator,
647
+ height=self.height,
648
+ width=self.width,
649
+ )
650
+ return result.images # list[PIL.Image]
651
+
652
+
653
+ def load_hunyuanimage21(device: torch.device):
654
+ from diffusers import HunyuanImagePipeline
655
+ MODEL_ID = "hunyuanvideo-community/HunyuanImage-2.1-Diffusers"
656
+ pipe = HunyuanImagePipeline.from_pretrained(MODEL_ID, dtype=torch.bfloat16).to(device)
657
+ print("[model] HunyuanImage-2.1 loaded successfully!")
658
+ return HunyuanImageWrapper(pipe)
659
+
660
+ # -----------------------------------------------------------------------------
661
+ # ============================ ADD NEW MODELS HERE ========================
662
+ # -----------------------------------------------------------------------------
663
+ # Template for registering a new text-to-image model:
664
+ #
665
+ # 1. Write a loader function `load_<your_model>(device) -> wrapper`.
666
+ # - If it's a standard diffusers pipeline, just load it and return
667
+ # DiffusersPipelineWrapper(pipe) — exactly like load_sd15 above.
668
+ # - If it's NOT a diffusers pipeline (custom repo, different API, etc.),
669
+ # write a small wrapper class with a matching method:
670
+ #
671
+ # def generate(self, prompt, num_inference_steps, guidance_scale,
672
+ # num_images_per_prompt, generator) -> list[PIL.Image]:
673
+ # ...
674
+ #
675
+ # `generator` will be a list of torch.Generator objects (one per
676
+ # image) or None — handle whichever your backend supports.
677
+ #
678
+ # 2. Register it in MODEL_LOADERS below with the --model_name string
679
+ # you want to use on the command line.
680
+ #
681
+ # Example skeleton:
682
+ #
683
+ # def load_my_new_model(device: torch.device):
684
+ # # from my_package import MyPipeline
685
+ # # pipe = MyPipeline.from_pretrained("org/my-model").to(device)
686
+ # # return DiffusersPipelineWrapper(pipe) # if diffusers-style
687
+ # raise NotImplementedError("Implement load_my_new_model() and register it below.")
688
+ #
689
+ # MODEL_LOADERS["my_new_model"] = load_my_new_model
690
+ # -----------------------------------------------------------------------------
691
+
692
+ # MODEL_LOADERS: Dict[str, Callable[[torch.device], Any]] = {
693
+ # "sd15": load_sd15,
694
+ # "sdxl": load_sdxl,
695
+ # "sana15": load_sana15,
696
+ # # "my_new_model": load_my_new_model, # <- uncomment once implemented above
697
+ # }
698
+
699
+ # MODEL_LOADERS: Dict[str, Callable[[torch.device], Any]] = {
700
+ # "sd15": load_sd15,
701
+ # "sdxl": load_sdxl,
702
+ # "sana15": load_sana15,
703
+ # "flux": load_flux,
704
+ # }
705
+ MODEL_LOADERS: Dict[str, Callable[[torch.device], Any]] = {
706
+ "sd15": load_sd15,
707
+ "sdxl": load_sdxl,
708
+ "sana15": load_sana15,
709
+ "flux": load_flux, # if you added FLUX.2-dev earlier
710
+ "flux_dev": load_flux_dev, # FLUX.1-dev
711
+ "qwen_image": load_qwen_image,
712
+ "sd35_medium": load_sd35_medium,
713
+ "sd35_large": load_sd35_large,
714
+ "pixart_alpha": load_pixart_alpha,
715
+ "auraflow": load_auraflow,
716
+ "kandinsky22": load_kandinsky22,
717
+ "pixart_sigma": load_pixart_sigma,
718
+ "z_image": load_z_image,
719
+ "hunyuanimage21": load_hunyuanimage21,
720
+ "kandinsky5": load_kandinsky5,
721
+ "kandinsky31": load_kandinsky31,
722
+ }
723
+
724
+
725
+ def unload_model(wrapper) -> None:
726
+ """Free GPU memory before loading the next model in a sweep."""
727
+ del wrapper
728
+ gc.collect()
729
+ if torch.cuda.is_available():
730
+ torch.cuda.empty_cache()
731
+
732
+
733
+ def generate_images(
734
+ records: List[PromptRecord],
735
+ wrapper,
736
+ root_output_dir: Path,
737
+ images_per_prompt: int,
738
+ num_inference_steps: int,
739
+ guidance_scale: float,
740
+ seed_base: Optional[int],
741
+ device: torch.device,
742
+ skip_existing: bool,
743
+ ) -> None:
744
+ """
745
+ Generates images for each unique (artist, artwork) pair in `records` and
746
+ saves them to the exact paths already computed in record.gen_paths — so
747
+ generation output and mapping.json can never disagree on filenames.
748
+ """
749
+ root_output_dir.mkdir(parents=True, exist_ok=True)
750
+
751
+ # Dedup by (artist, artwork): identical pairs share identical gen_paths,
752
+ # so there's no reason to render them twice (mirrors ET_infer.py's
753
+ # prompt-level dedup, but keyed on the actual output-path identity).
754
+ seen = set()
755
+ unique_records: List[PromptRecord] = []
756
+ for r in records:
757
+ key = (r.artist, r.artwork)
758
+ if key in seen:
759
+ continue
760
+ seen.add(key)
761
+ unique_records.append(r)
762
+
763
+ total = len(unique_records)
764
+ for idx, rec in enumerate(unique_records, start=1):
765
+ if skip_existing and all(Path(p).exists() for p in rec.gen_paths):
766
+ print(f"[{idx}/{total}] [skip] already generated: '{rec.artwork}' x '{rec.artist}'")
767
+ continue
768
+
769
+ print(f"\n[{idx}/{total}] Generating: '{rec.artwork}' in the style of '{rec.artist}'")
770
+ print(f"Prompt: {rec.prompt}")
771
+
772
+ generator = None
773
+ if seed_base is not None:
774
+ generator = [
775
+ torch.Generator(device=device).manual_seed(seed_base + idx * 1000 + k)
776
+ for k in range(images_per_prompt)
777
+ ]
778
+
779
+ images = wrapper.generate(
780
+ prompt=rec.prompt,
781
+ num_inference_steps=num_inference_steps,
782
+ guidance_scale=guidance_scale,
783
+ num_images_per_prompt=images_per_prompt,
784
+ generator=generator,
785
+ )
786
+
787
+ for k, (image, out_path) in enumerate(zip(images, rec.gen_paths), start=1):
788
+ Path(out_path).parent.mkdir(parents=True, exist_ok=True)
789
+ image.save(out_path)
790
+ print(f" -> Saved: {out_path}")
791
+
792
+
793
+ # =============================================================================
794
+ # STAGE 3 — proximity scoring (clip / csd / lpips) (was ET_eval_fixed.py)
795
+ # =============================================================================
796
+
797
+ @dataclass
798
+ class MetricHandlers:
799
+ load_embedding: Callable[[str], torch.Tensor]
800
+ compute_score: Callable[[torch.Tensor, torch.Tensor], float]
801
+ score_column_name: str
802
+ sort_ascending: bool
803
+ cache_embeddings: bool
804
+ model: Any = field(default=None, repr=False) # kept alive for the closures' lifetime
805
+
806
+
807
+ def amp_context(device: torch.device, use_amp: bool):
808
+ if use_amp and device.type == "cuda":
809
+ return torch.amp.autocast("cuda", dtype=torch.float16)
810
+ return contextlib.nullcontext()
811
+
812
+
813
+ def build_clip_handlers(device: torch.device, use_amp: bool) -> MetricHandlers:
814
+ from transformers import CLIPModel, CLIPProcessor
815
+ CLIP_MODEL_ID = "openai/clip-vit-base-patch32"
816
+ model = CLIPModel.from_pretrained(CLIP_MODEL_ID).to(device)
817
+ processor = CLIPProcessor.from_pretrained(CLIP_MODEL_ID)
818
+ model.eval()
819
+ print(f"[metric] CLIP model '{CLIP_MODEL_ID}' loaded successfully!")
820
+
821
+ def load_embedding(image_path: str) -> torch.Tensor:
822
+ with Image.open(image_path) as im:
823
+ im = im.convert("RGB")
824
+ inputs = processor(images=im, return_tensors="pt")
825
+ pixel_values = inputs["pixel_values"].to(device)
826
+ with torch.no_grad(), amp_context(device, use_amp):
827
+ vision_outputs = model.vision_model(pixel_values=pixel_values)
828
+ pooled_output = vision_outputs.pooler_output
829
+ img_feat = model.visual_projection(pooled_output)
830
+ img_feat = img_feat / (img_feat.norm(p=2, dim=-1, keepdim=True) + 1e-8)
831
+ return img_feat
832
+
833
+ def compute_score(a: torch.Tensor, b: torch.Tensor) -> float:
834
+ a = a / (a.norm(p=2, dim=-1, keepdim=True) + 1e-8)
835
+ b = b / (b.norm(p=2, dim=-1, keepdim=True) + 1e-8)
836
+ return float((a @ b.T).item())
837
+
838
+ return MetricHandlers(load_embedding, compute_score, "clip_cosine_similarity",
839
+ sort_ascending=False, cache_embeddings=True, model=model)
840
+
841
+
842
+ def build_csd_handlers(
843
+ device: torch.device, use_amp: bool, csd_repo_dir: Path, model_path: Path,
844
+ use_cosine_similarity: bool,
845
+ ) -> MetricHandlers:
846
+ from torchvision import transforms
847
+ import torchvision.transforms.functional as TF
848
+
849
+ if not csd_repo_dir.is_dir():
850
+ sys.exit(f"[ERROR] --csd_repo_dir not found: {csd_repo_dir}")
851
+ if not model_path.is_file():
852
+ sys.exit(f"[ERROR] --csd_model_path not found: {model_path}")
853
+
854
+ csd_repo_abs = csd_repo_dir.resolve()
855
+ csd_parent = csd_repo_abs.parent
856
+ csd_pkg_name = csd_repo_abs.name
857
+
858
+ if str(csd_parent) not in sys.path:
859
+ sys.path.insert(0, str(csd_parent))
860
+
861
+ init_file = csd_repo_abs / "__init__.py"
862
+ if not init_file.exists():
863
+ init_file.touch()
864
+ print(f"[metric] Created missing {init_file} to enable package-relative imports.")
865
+
866
+ csd_model_mod = importlib.import_module(f"{csd_pkg_name}.model")
867
+ CSD_CLIP = csd_model_mod.CSD_CLIP
868
+
869
+ if hasattr(csd_model_mod, "convert_state_dict"):
870
+ convert_state_dict = csd_model_mod.convert_state_dict
871
+ else:
872
+ csd_utils_mod = importlib.import_module(f"{csd_pkg_name}.utils")
873
+ convert_state_dict = csd_utils_mod.convert_state_dict
874
+
875
+ model = CSD_CLIP("vit_large", "default")
876
+ checkpoint = torch.load(str(model_path), map_location="cpu", weights_only=False)
877
+ state_dict = convert_state_dict(checkpoint["model_state_dict"])
878
+ model.load_state_dict(state_dict, strict=False)
879
+ model = model.to(device)
880
+ model.eval()
881
+
882
+ normalize = transforms.Normalize(
883
+ (0.48145466, 0.4578275, 0.40821073),
884
+ (0.26862954, 0.26130258, 0.27577711),
885
+ )
886
+ preprocess = transforms.Compose([
887
+ transforms.Resize(size=224, interpolation=TF.InterpolationMode.BICUBIC),
888
+ transforms.CenterCrop(224),
889
+ transforms.ToTensor(),
890
+ normalize,
891
+ ])
892
+ print(f"[metric] CSD model loaded from '{model_path}' | cosine_sim={use_cosine_similarity}")
893
+
894
+ def load_embedding(image_path: str) -> torch.Tensor:
895
+ with Image.open(image_path) as im:
896
+ im = im.convert("RGB")
897
+ img_tensor = preprocess(im).unsqueeze(0).to(device)
898
+ with amp_context(device, use_amp):
899
+ output = model(img_tensor)
900
+ style_output = output[1] if isinstance(output, (tuple, list)) else output
901
+ return style_output
902
+
903
+ def compute_score(a: torch.Tensor, b: torch.Tensor) -> float:
904
+ if use_cosine_similarity:
905
+ eps = 1e-8
906
+ a = a / (a.norm(dim=1, keepdim=True) + eps)
907
+ b = b / (b.norm(dim=1, keepdim=True) + eps)
908
+ return float((a @ b.T).item())
909
+
910
+ return MetricHandlers(load_embedding, compute_score, "csd_score",
911
+ sort_ascending=False, cache_embeddings=True, model=model)
912
+
913
+
914
+ def build_lpips_handlers(device: torch.device) -> MetricHandlers:
915
+ import torch.nn.functional as F
916
+ import torchvision.transforms as T
917
+ from PIL import ImageOps
918
+
919
+ try:
920
+ import lpips
921
+ except ImportError:
922
+ raise RuntimeError("lpips not installed. Run: pip install lpips")
923
+
924
+ LPIPS_NET = "alex"
925
+ INTERP_MODE = "bicubic"
926
+ to_tensor_01 = T.Compose([T.ToTensor()])
927
+
928
+ # -------------------------------------------------------------------
929
+ # Load AlexNet's ImageNet weights from a LOCAL FILE, bypassing the
930
+ # network download entirely.
931
+ #
932
+ # Same underlying mechanism as the working Welkin reference script
933
+ # (also applied to MotifDuels.py's build_lpips_tournament_metric):
934
+ # lpips.LPIPS(net="alex") calls torchvision.models.alexnet
935
+ # (pretrained=True), which always resolves its weights via
936
+ # torch.hub -- checking <TORCH_HOME>/hub/checkpoints/<filename>.pth
937
+ # BEFORE ever attempting a download. So instead of intercepting the
938
+ # download call, we seed that exact cache location with the local
939
+ # file first; torch.hub then just finds it there and never reaches
940
+ # the network.
941
+ #
942
+ # Set this to your local .pth file.
943
+ ALEXNET_WEIGHTS_LOCAL_PATH = "/workspace/data/Ninad/Style_Leakage/Large_Experiments/Imitation_on_all/Art_FID/alexnet-owt-7be5be79.pth"
944
+
945
+ def _prepare_torchvision_cache_for_alexnet(local_path: str) -> str:
946
+ """Ensure torchvision finds alexnet-owt-7be5be79.pth offline by
947
+ placing it in the hub checkpoints cache. Returns the destination path."""
948
+ import os
949
+ import shutil
950
+
951
+ if not os.path.isfile(local_path):
952
+ raise FileNotFoundError(f"Local AlexNet weights not found at: {local_path}")
953
+
954
+ torch_home = os.environ.get("TORCH_HOME", os.path.join(Path.home(), ".cache", "torch"))
955
+ checkpoints_dir = os.path.join(torch_home, "hub", "checkpoints")
956
+ os.makedirs(checkpoints_dir, exist_ok=True)
957
+
958
+ dest = os.path.join(checkpoints_dir, "alexnet-owt-7be5be79.pth")
959
+
960
+ need_copy = True
961
+ if os.path.exists(dest):
962
+ try:
963
+ need_copy = os.path.getsize(local_path) != os.path.getsize(dest)
964
+ except Exception:
965
+ need_copy = True
966
+
967
+ if need_copy:
968
+ shutil.copy2(local_path, dest)
969
+ print(f"[CACHE] Seeded TorchVision cache: {dest}")
970
+ else:
971
+ print(f"[CACHE] TorchVision cache already contains AlexNet weights: {dest}")
972
+
973
+ return dest
974
+
975
+ if ALEXNET_WEIGHTS_LOCAL_PATH:
976
+ try:
977
+ _prepare_torchvision_cache_for_alexnet(ALEXNET_WEIGHTS_LOCAL_PATH)
978
+ except FileNotFoundError as e:
979
+ print(f"[WARN] {e} -- falling back to normal download.")
980
+
981
+ loss_fn = lpips.LPIPS(net=LPIPS_NET).to(device)
982
+ loss_fn.eval()
983
+ print(f"[metric] LPIPS model (net={LPIPS_NET}) loaded successfully!")
984
+
985
+ def load_embedding(image_path: str) -> torch.Tensor:
986
+ with Image.open(image_path) as im:
987
+ im = ImageOps.exif_transpose(im).convert("RGB")
988
+ x = to_tensor_01(im) * 2.0 - 1.0 # [-1, 1]
989
+ return x.unsqueeze(0).to(device)
990
+
991
+ def compute_score(a: torch.Tensor, b: torch.Tensor) -> float:
992
+ if a.shape[2:] != b.shape[2:]:
993
+ h_t, w_t = b.shape[2], b.shape[3]
994
+ try:
995
+ a = F.interpolate(a, size=(h_t, w_t), mode=INTERP_MODE, align_corners=False, antialias=True)
996
+ except TypeError:
997
+ a = F.interpolate(a, size=(h_t, w_t), mode=INTERP_MODE, align_corners=False)
998
+ return float(loss_fn(a, b).item())
999
+
1000
+ return MetricHandlers(load_embedding, compute_score, "lpips_distance",
1001
+ sort_ascending=True, cache_embeddings=False, model=loss_fn)
1002
+
1003
+
1004
+ METRIC_BUILDERS = {"clip": build_clip_handlers, "csd": build_csd_handlers, "lpips": build_lpips_handlers}
1005
+
1006
+
1007
+ def is_nonempty_str(s) -> bool:
1008
+ return isinstance(s, str) and len(s.strip()) > 0
1009
+
1010
+
1011
+ def has_valid_image_extension(path: str) -> bool:
1012
+ return Path(path).suffix.lower() in VALID_IMAGE_EXTS
1013
+
1014
+
1015
+ def short_hash(text: str, length: int = 8) -> str:
1016
+ return hashlib.sha1(text.encode("utf-8")).hexdigest()[:length]
1017
+
1018
+
1019
+ def unique_target_path(dst_dir: Path, base_name: str) -> Path:
1020
+ target = dst_dir / base_name
1021
+ if not target.exists():
1022
+ return target
1023
+ stem, ext = Path(base_name).stem, Path(base_name).suffix
1024
+ i = 1
1025
+ while True:
1026
+ candidate = dst_dir / f"{stem}__{i}{ext}"
1027
+ if not candidate.exists():
1028
+ return candidate
1029
+ i += 1
1030
+
1031
+
1032
+ def copy_with_rank_and_hash(src_path: str, dst_dir: Path, rank: int) -> Path:
1033
+ dst_dir.mkdir(parents=True, exist_ok=True)
1034
+ base = Path(src_path).name
1035
+ h = short_hash(src_path)
1036
+ # ranked_name = f"[rank:{rank:02d}]__{h}__{base}"
1037
+ ranked_name = f"[rank-{rank:02d}]__{h}__{base}"
1038
+ target_path = unique_target_path(dst_dir, ranked_name)
1039
+ shutil.copy2(src_path, target_path)
1040
+ return target_path
1041
+
1042
+
1043
+ def save_top_pairs(results: List[Tuple[str, List[str], float]], top_root: Path, top_n: int) -> None:
1044
+ gen_dir = top_root / f"top{top_n}_generated"
1045
+ orig_dir = top_root / f"top{top_n}_original"
1046
+ print(f"[metric] Saving top-{top_n} pairs -> generated: {gen_dir} | original: {orig_dir}")
1047
+ for rank, (orig_path, gen_paths, score) in enumerate(results[:top_n], start=1):
1048
+ try:
1049
+ orig_dst = copy_with_rank_and_hash(orig_path, orig_dir, rank)
1050
+ gen_dsts = [copy_with_rank_and_hash(gp, gen_dir, rank) for gp in gen_paths]
1051
+ print(f"[TOP {rank:02d}] avg_score={score:.6f} orig: {orig_dst} gen: {gen_dsts}")
1052
+ except Exception as e:
1053
+ print(f"[ERROR] Failed copying pair rank={rank} ({orig_path}): {e}")
1054
+
1055
+
1056
+ def score_mapping(
1057
+ mapping: Dict[str, List[str]],
1058
+ handlers: MetricHandlers,
1059
+ output_csv: Path,
1060
+ top_save_root: Optional[Path],
1061
+ top_n: int,
1062
+ ) -> None:
1063
+ t0 = time.time()
1064
+ orig_embeddings: Dict[str, torch.Tensor] = {}
1065
+ gen_embeddings: Dict[str, torch.Tensor] = {}
1066
+ results: List[Tuple[str, List[str], float]] = []
1067
+ errors: List[str] = []
1068
+
1069
+ total = len(mapping)
1070
+ print(f"[metric] Processing {total} originals...")
1071
+
1072
+ with torch.no_grad():
1073
+ for idx, (orig_path, gen_paths) in enumerate(mapping.items(), start=1):
1074
+ if not Path(orig_path).is_file():
1075
+ msg = f"Missing original file: {orig_path}"
1076
+ print(f"[WARN] {msg}"); errors.append(msg); continue
1077
+
1078
+ try:
1079
+ if handlers.cache_embeddings:
1080
+ if orig_path not in orig_embeddings:
1081
+ orig_embeddings[orig_path] = handlers.load_embedding(orig_path).detach().cpu()
1082
+ a = orig_embeddings[orig_path]
1083
+ else:
1084
+ a = handlers.load_embedding(orig_path)
1085
+ except Exception as e:
1086
+ msg = f"Error embedding original ({orig_path}): {e}"
1087
+ print(f"[ERROR] {msg}"); errors.append(msg); continue
1088
+
1089
+ per_gen_scores: List[float] = []
1090
+ for gen_path in gen_paths:
1091
+ if not Path(gen_path).is_file():
1092
+ msg = f"Missing generated file: {gen_path} (original={orig_path})"
1093
+ print(f"[WARN] {msg}"); errors.append(msg); continue
1094
+ try:
1095
+ if handlers.cache_embeddings:
1096
+ if gen_path not in gen_embeddings:
1097
+ gen_embeddings[gen_path] = handlers.load_embedding(gen_path).detach().cpu()
1098
+ b = gen_embeddings[gen_path]
1099
+ else:
1100
+ b = handlers.load_embedding(gen_path)
1101
+ per_gen_scores.append(handlers.compute_score(a, b))
1102
+ except Exception as e:
1103
+ msg = f"Error scoring pair ({orig_path}, {gen_path}): {e}"
1104
+ print(f"[ERROR] {msg}"); errors.append(msg)
1105
+
1106
+ if not per_gen_scores:
1107
+ msg = f"No valid generated scores for original={orig_path}; skipping."
1108
+ print(f"[WARN] {msg}"); errors.append(msg); continue
1109
+
1110
+ avg_score = sum(per_gen_scores) / len(per_gen_scores)
1111
+ results.append((orig_path, gen_paths, avg_score))
1112
+
1113
+ if idx % 50 == 0 or idx == total:
1114
+ print(f"[metric] {idx}/{total} processed | avg_score={avg_score:.6f}")
1115
+
1116
+ results.sort(key=lambda x: x[2], reverse=not handlers.sort_ascending)
1117
+
1118
+ output_csv.parent.mkdir(parents=True, exist_ok=True)
1119
+ with open(output_csv, "w", newline="", encoding="utf-8") as f:
1120
+ writer = csv.writer(f)
1121
+ writer.writerow(["rank", "original_path", "generated_paths", "num_generated", handlers.score_column_name])
1122
+ for rank, (orig_path, gen_paths, avg_score) in enumerate(results, start=1):
1123
+ writer.writerow([rank, orig_path, "|".join(gen_paths), len(gen_paths), f"{avg_score:.8f}"])
1124
+
1125
+ print(f"[DONE] {len(results)} rows -> {output_csv} | Elapsed: {time.time()-t0:.2f}s | Errors: {len(errors)}")
1126
+ if errors:
1127
+ err_log = output_csv.with_name(output_csv.stem + "_errors.log")
1128
+ err_log.write_text("\n".join(errors), encoding="utf-8")
1129
+ print(f"[metric] Error log: {err_log}")
1130
+
1131
+ if top_save_root is not None:
1132
+ save_top_pairs(results, top_save_root, top_n)
1133
+
1134
+
1135
+ def unload_metric(handlers: MetricHandlers) -> None:
1136
+ del handlers
1137
+ gc.collect()
1138
+ if torch.cuda.is_available():
1139
+ torch.cuda.empty_cache()
1140
+
1141
+
1142
+ # =============================================================================
1143
+ # Orchestration
1144
+ # =============================================================================
1145
+
1146
+ def parse_args() -> argparse.Namespace:
1147
+ p = argparse.ArgumentParser(
1148
+ description="Self-contained ArtArena Entry-Trials pipeline: prep + generate + score.",
1149
+ formatter_class=argparse.ArgumentDefaultsHelpFormatter,
1150
+ )
1151
+
1152
+ p.add_argument("--list_models", action="store_true",
1153
+ help="Print registered model names and exit.")
1154
+
1155
+ p.add_argument("--datasets", nargs="+",
1156
+ help="One or more dataset directories.")
1157
+ p.add_argument("--models", nargs="+", default=["sd15"],
1158
+ help=f"One or more model names. Registered: {list(MODEL_LOADERS.keys())}")
1159
+ p.add_argument("--metrics", nargs="+", default=["clip"], choices=list(METRIC_BUILDERS.keys()),
1160
+ help="One or more proximity metrics.")
1161
+ p.add_argument("--output_root",
1162
+ help="Root directory for all outputs.")
1163
+
1164
+ # prep
1165
+ p.add_argument("--num_imgs", type=int, default=1,
1166
+ help="Images per prompt (used for both dataset prep and generation).")
1167
+ p.add_argument("--name_split", choices=["first", "last"], default="last",
1168
+ help="Fallback-only: used just for files the frequency-based parser can't resolve.")
1169
+ p.add_argument("--min_artist_count", type=int, default=3,
1170
+ help="Min files an artist must recur in for the frequency-based parser to trust it. "
1171
+ "Lower this if you have artists with very few paintings.")
1172
+ p.add_argument("--max_artist_tokens", type=int, default=4,
1173
+ help="Max artist-name length in underscore-tokens to consider (e.g. 'Vincent_van_Gogh' = 3).")
1174
+ p.add_argument("--artist_artwork_mapping", default=None,
1175
+ help="Path to a pre-built artist_artwork_mapping.json to use for EVERY dataset in this "
1176
+ "run, instead of the per-dataset auto path "
1177
+ "(<output_root>/<dataset_name>/artist_artwork_mapping.json). Same format this "
1178
+ "script already writes/reuses automatically. Build it once with a single prep-only "
1179
+ "run, then pass it explicitly to every parallel --start_index/--end_index job so "
1180
+ "they all resolve identical artist/artwork pairs instead of racing to build/write "
1181
+ "it independently.")
1182
+
1183
+ # infer chunking (for running several instances of this script in parallel)
1184
+ p.add_argument("--start_index", type=int, default=0,
1185
+ help="Only generate records[start_index:end_index] for each dataset (0-based, inclusive). "
1186
+ "prompts.pt/mapping.json always cover the FULL dataset regardless of this range — "
1187
+ "only the image-generation loop is sliced. Use this to split one large dataset's "
1188
+ "generation across multiple parallel script instances.")
1189
+ p.add_argument("--end_index", type=int, default=None,
1190
+ help="Exclusive upper bound for the generation slice (see --start_index). "
1191
+ "Omit to generate through the end of the dataset's record list.")
1192
+
1193
+ # infer
1194
+ p.add_argument("--num_inference_steps", type=int, default=25)
1195
+ p.add_argument("--guidance_scale", type=float, default=7.5)
1196
+ p.add_argument("--seed_base", type=int, default=42,
1197
+ help="Pass -1 for fully random generation.")
1198
+
1199
+ # eval
1200
+ p.add_argument("--top_n", type=int, default=20)
1201
+ p.add_argument("--no_amp", action="store_true", help="Disable AMP during scoring.")
1202
+ p.add_argument("--use_cosine_similarity_csd", action="store_true",
1203
+ help="[CSD only] L2-normalise before dot product.")
1204
+ p.add_argument("--csd_repo_dir", default=str(SCRIPT_DIR / "CSD"),
1205
+ help="Folder containing CSD's model.py/utils.py.")
1206
+ p.add_argument("--csd_model_path", default=str(SCRIPT_DIR / "CSD" / "checkpoint.pth"),
1207
+ help="Path to the CSD checkpoint .pth file.")
1208
+
1209
+ # orchestration
1210
+ p.add_argument("--stages", nargs="+", default=["prep", "infer", "eval"],
1211
+ choices=["prep", "infer", "eval"],
1212
+ help="Which stages to run. 'prep' is cheap and always safe to include; "
1213
+ "excluding it requires prompts.pt/mapping.json to already exist.")
1214
+ p.add_argument("--skip_existing", action="store_true",
1215
+ help="Skip generation/scoring for outputs that already exist (resumable sweeps).")
1216
+ p.add_argument("--dry_run", action="store_true",
1217
+ help="Print the planned dataset x model x metric grid and exit, without loading anything.")
1218
+
1219
+ args = p.parse_args()
1220
+
1221
+ if args.list_models:
1222
+ return args
1223
+ if not args.datasets or not args.output_root:
1224
+ p.error("--datasets and --output_root are required (unless using --list_models)")
1225
+ for m in args.models:
1226
+ if m not in MODEL_LOADERS:
1227
+ p.error(f"Unknown model '{m}'. Registered: {list(MODEL_LOADERS.keys())}")
1228
+ if args.start_index < 0:
1229
+ p.error("--start_index must be >= 0")
1230
+ if args.end_index is not None and args.end_index <= args.start_index:
1231
+ p.error(f"--end_index ({args.end_index}) must be greater than --start_index ({args.start_index})")
1232
+ return args
1233
+
1234
+
1235
+ def main() -> None:
1236
+ args = parse_args()
1237
+
1238
+ if args.list_models:
1239
+ print("Registered models:")
1240
+ for name in MODEL_LOADERS:
1241
+ print(f" - {name}")
1242
+ return
1243
+
1244
+ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
1245
+ if device.type == "cpu":
1246
+ print("[WARN] CUDA not available - running on CPU. This will be very slow.")
1247
+
1248
+ output_root = Path(args.output_root).resolve()
1249
+ output_root.mkdir(parents=True, exist_ok=True)
1250
+ dataset_dirs = [Path(d).resolve() for d in args.datasets]
1251
+ use_amp = not args.no_amp
1252
+
1253
+ print(
1254
+ f"[i] Sweep: {len(dataset_dirs)} dataset(s) x {len(args.models)} model(s) "
1255
+ f"x {len(args.metrics)} metric(s) | stages={args.stages}"
1256
+ )
1257
+
1258
+ def combo_dir(dataset_dir: Path, model_name: str) -> Path:
1259
+ return output_root / dataset_dir.name / model_name
1260
+
1261
+ if args.dry_run:
1262
+ for dataset_dir in dataset_dirs:
1263
+ for model_name in args.models:
1264
+ cd = combo_dir(dataset_dir, model_name)
1265
+ print(f"[dry_run] {dataset_dir.name} x {model_name} -> {cd}")
1266
+ for metric in args.metrics:
1267
+ print(f"[dry_run] eval[{metric}] -> {cd / 'eval' / metric}")
1268
+ return
1269
+
1270
+ failures: List[str] = []
1271
+ t_start = time.time()
1272
+
1273
+ # ---------------- Phase 1: prep + generation (model-outer loop) -----------------
1274
+ for model_name in args.models:
1275
+ wrapper = None
1276
+ if "infer" in args.stages:
1277
+ print(f"\n{'='*80}\n[Loading model] {model_name}\n{'='*80}")
1278
+ try:
1279
+ wrapper = MODEL_LOADERS[model_name](device)
1280
+ except Exception as e:
1281
+ msg = f"Failed to load model '{model_name}': {e}"
1282
+ print(f"[ERROR] {msg}")
1283
+ failures.append(msg)
1284
+ # continue
1285
+
1286
+ for dataset_dir in dataset_dirs:
1287
+ cd = combo_dir(dataset_dir, model_name)
1288
+ gen_dir = cd / "generated"
1289
+ print(f"\n--- dataset={dataset_dir.name} | model={model_name} ---")
1290
+
1291
+ records: Optional[List[PromptRecord]] = None
1292
+ if "prep" in args.stages or "infer" in args.stages:
1293
+ try:
1294
+ mapping_path = (
1295
+ Path(args.artist_artwork_mapping) if args.artist_artwork_mapping
1296
+ else (output_root / dataset_dir.name / "artist_artwork_mapping.json")
1297
+ )
1298
+ if "prep" in args.stages:
1299
+ # This is the "producer" run: build fresh from the dataset scan
1300
+ # (reusing mapping_path's artist/artwork mapping if it already
1301
+ # exists, exactly as before) and ALWAYS (over)save the FULL
1302
+ # record set — prompts.pt/mapping.json must stay complete
1303
+ # regardless of --start_index/--end_index, since the eval
1304
+ # stage later scores against the whole dataset once all
1305
+ # parallel generation jobs have finished their chunks.
1306
+ records = build_prompts_and_mapping(
1307
+ dataset_dir, gen_dir, args.num_imgs, args.name_split,
1308
+ artist_artwork_mapping_path=mapping_path,
1309
+ min_artist_count=args.min_artist_count,
1310
+ max_artist_tokens=args.max_artist_tokens,
1311
+ )
1312
+ save_prompts_pt(records, cd / "prompts.pt")
1313
+ save_mapping_json(records, cd / "mapping.json")
1314
+ else:
1315
+ # 'prep' excluded: this is a "consumer" run (e.g. a parallel
1316
+ # --start_index/--end_index chunk job) -- reuse whatever the
1317
+ # producer run already saved. Never rebuilds, never rewrites,
1318
+ # so no race with sibling jobs running at the same time.
1319
+ artist_artwork_map = (
1320
+ load_artist_artwork_mapping(mapping_path) if mapping_path.is_file() else None
1321
+ )
1322
+ records = load_existing_records(
1323
+ cd / "prompts.pt", cd / "mapping.json", args.name_split,
1324
+ artist_artwork_mapping=artist_artwork_map,
1325
+ )
1326
+ except (SystemExit, FileNotFoundError, ValueError) as e:
1327
+ msg = f"prep FAILED for dataset={dataset_dir.name}: {e}"
1328
+ print(f"[ERROR] {msg}")
1329
+ failures.append(msg)
1330
+ continue
1331
+
1332
+ if "infer" in args.stages and wrapper is not None and records is not None:
1333
+ start = min(args.start_index, len(records))
1334
+ end = len(records) if args.end_index is None else min(args.end_index, len(records))
1335
+ chunk = records[start:end]
1336
+ print(f"[i] Generation slice: records[{start}:{end}] of {len(records)} total "
1337
+ f"({len(chunk)} record(s) in this run)")
1338
+ if not chunk:
1339
+ print(f"[WARN] Empty slice for dataset={dataset_dir.name} "
1340
+ f"(start_index={args.start_index} >= {len(records)} total records) — nothing to generate.")
1341
+ else:
1342
+ try:
1343
+ generate_images(
1344
+ chunk, wrapper, gen_dir,
1345
+ images_per_prompt=args.num_imgs,
1346
+ num_inference_steps=args.num_inference_steps,
1347
+ guidance_scale=args.guidance_scale,
1348
+ seed_base=(args.seed_base if args.seed_base >= 0 else None),
1349
+ device=device,
1350
+ skip_existing=args.skip_existing,
1351
+ )
1352
+ except Exception as e:
1353
+ msg = f"infer FAILED for dataset={dataset_dir.name} model={model_name}: {e}"
1354
+ print(f"[ERROR] {msg}")
1355
+ failures.append(msg)
1356
+
1357
+ if wrapper is not None:
1358
+ unload_model(wrapper)
1359
+
1360
+ # ---------------- Phase 2: evaluation (metric-outer loop) -----------------
1361
+ if "eval" in args.stages:
1362
+ for metric in args.metrics:
1363
+ print(f"\n{'='*80}\n[Loading metric] {metric}\n{'='*80}")
1364
+ try:
1365
+ if metric == "clip":
1366
+ handlers = build_clip_handlers(device, use_amp)
1367
+ elif metric == "lpips":
1368
+ handlers = build_lpips_handlers(device)
1369
+ elif metric == "csd":
1370
+ handlers = build_csd_handlers(
1371
+ device, use_amp,
1372
+ csd_repo_dir=Path(args.csd_repo_dir),
1373
+ model_path=Path(args.csd_model_path),
1374
+ use_cosine_similarity=args.use_cosine_similarity_csd,
1375
+ )
1376
+ else:
1377
+ raise ValueError(metric)
1378
+ except Exception as e:
1379
+ msg = f"Failed to load metric '{metric}': {e}"
1380
+ print(f"[ERROR] {msg}")
1381
+ failures.append(msg)
1382
+ continue
1383
+
1384
+ for model_name in args.models:
1385
+ for dataset_dir in dataset_dirs:
1386
+ cd = combo_dir(dataset_dir, model_name)
1387
+ mapping_json = cd / "mapping.json"
1388
+ if not mapping_json.exists():
1389
+ msg = f"eval SKIPPED (no mapping.json) for dataset={dataset_dir.name} model={model_name}"
1390
+ print(f"[WARN] {msg}")
1391
+ failures.append(msg)
1392
+ continue
1393
+
1394
+ eval_dir = cd / "eval" / metric
1395
+ scores_csv = eval_dir / f"{metric}_scores.csv"
1396
+ if args.skip_existing and scores_csv.exists():
1397
+ print(f"[skip] {scores_csv} already exists")
1398
+ continue
1399
+
1400
+ print(f"\n--- eval dataset={dataset_dir.name} | model={model_name} | metric={metric} ---")
1401
+ with open(mapping_json, "r", encoding="utf-8") as f:
1402
+ mapping = json.load(f)
1403
+ mapping = {
1404
+ o: (g if isinstance(g, list) else [g])
1405
+ for o, g in mapping.items()
1406
+ if is_nonempty_str(o) and has_valid_image_extension(o)
1407
+ }
1408
+
1409
+ try:
1410
+ score_mapping(mapping, handlers, scores_csv, eval_dir / "candidates", args.top_n)
1411
+ except Exception as e:
1412
+ msg = f"eval FAILED for dataset={dataset_dir.name} model={model_name} metric={metric}: {e}"
1413
+ print(f"[ERROR] {msg}")
1414
+ failures.append(msg)
1415
+
1416
+ unload_metric(handlers)
1417
+
1418
+ elapsed = time.time() - t_start
1419
+ print(f"\n{'='*80}\n[Done] Pipeline finished in {elapsed:.1f}s")
1420
+ if failures:
1421
+ print(f"[!] {len(failures)} issue(s):")
1422
+ for f in failures:
1423
+ print(f" - {f}")
1424
+ sys.exit(1)
1425
+ print("[OK] All stages completed successfully.")
1426
+
1427
+
1428
+ if __name__ == "__main__":
1429
+ main()