isaac/isaac-image-scoring

Fork 0

feat: unified rank, rich centroid naming, siphon command #4

Merged

isaac merged 47 commits from unify-rank-siphon into main

2026-05-21 03:59:36 +00:00

lily commented

2026-05-21 03:57:38 +00:00

Contributor

Summary

Unified rank + siphon feature branch for aesthetic_scorer.py.

New features

Unified rank_folder — single function, all 3 modes auto-detected
Rich centroid naming — extract --name → siglip2_2026-05-20_name.pt with metadata
siphon N target — copy top N%, manifest CSV, --dry-run/--flatten/--move
list command — simple filename listing
--name for rank — matches by filename with extension
Percentile stretch — text-only and combined modes stretch clustered text scores from [p5, p95] → [0.1, 0.9], spreading clustered scores wider

Bug fixes

extract_features duplicate batch enqueue (last_enqueued_end guard)
extract_features empty batch pbar update
text-only rank crashes on pos-only/neg-only prompts
list command path check (skipped for list/test)
Dead code removed

Testing

python aesthetic_scorer.py test passes

## Summary Unified rank + siphon feature branch for aesthetic_scorer.py. ### New features - **Unified rank_folder** — single function, all 3 modes auto-detected - **Rich centroid naming** — extract --name → siglip2_2026-05-20_name.pt with metadata - **siphon N target** — copy top N%, manifest CSV, --dry-run/--flatten/--move - **list command** — simple filename listing - **--name for rank** — matches by filename with extension - **Percentile stretch** — text-only and combined modes stretch clustered text scores from [p5, p95] → [0.1, 0.9], spreading clustered scores wider ### Bug fixes - extract_features duplicate batch enqueue (last_enqueued_end guard) - extract_features empty batch pbar update - text-only rank crashes on pos-only/neg-only prompts - list command path check (skipped for list/test) - Dead code removed ### Testing python aesthetic_scorer.py test passes

lily added 48 commits

2026-05-21 03:57:38 +00:00

lkuijh 10a5cc51a1

Merge branch 'main' of https://forgejo.aizaku.ca/isaac/isaac-image-scoring 682336ba1f

feat: complete SigLIP2-only architecture — single model, single pass 296036cd2a

- Replace C-RADIO with SigLIP2 ViT-gopt-16-SigLIP2-384 for all commands
- rank-text: one embed per image, taste cosine + text cosine averaged raw
- extract: produces models/siglip_centroid.pt (1024-dim SigLIP2 centroid)
- Tokenizer: padding=max_length, max_length=64, explicit .lower() on all text
- Breaking: remove C-RADIO extractor, percentile rank, --resolution flag

Co-Authored-By: Claude <noreply@anthropic.com>

fix: SigLIP2 tokenizer test — use correct API (HFTokenizer returns Tensor) 83c493a856

fix: score_with_text uses batch matmul O(N) instead of per-image O(N*M) loop 7375868442

fix: parallelize file renames via ThreadPoolExecutor (8 workers) 55b0579eed

perf: enable cuDNN benchmark mode on CUDA for fixed-shape ViT inference 867ba018c5

Merge origin/main into feature/siglip-only 084ae86c57

fix: apply batch matmul to rank_text — O(B) not O(B*M) loop 8837193c5e

rank_text had the same per-image loop that score_with_text had.
Now both use:
  pos_sim = img_feats @ pos_embeds.T   # (B, M) matmul
  avg_pos = pos_sim.mean(dim=1)          # O(B) not O(B*M)

perf: letterbox resize to 384x384 before SigLIP2 encode (preserve aspect ratio) c4a06620c7

fix: remove dead OOM retry block; eliminate redundant intermediate variable 6a543eaba2

fix: FEATURE_DIM=1024→1536, open_clip ViT-gopt-16-SigLIP2-384 outputs 1536-dim 67d33404be

fix: encode_texts uses tokenizer() without explicit padding kwarg (HFTokenizer doesn't support it) d1e97334c7

good de141644d7

Merge branch 'feature/siglip-only' of https://forgejo.aizaku.ca/isaac/isaac-image-scoring into feature/siglip-only a3e6aa5694

perf: prefetch next batch disk reads while GPU processes current batch a60ada32cf

Merge remote-tracking branch 'origin/feature/siglip-only' into feature/siglip-only 30f6a85884

chore: fix stale comments — 1536-dim centroid (not 1024), remove misleading HEIC note b6a5f83d82

fix: close pbar and shutdown executor on exception in all three GPU functions ac8293e101

Pre-landing review found critical resource leaks:
- extract_features: pbar and executor leaked if encode_images raised
- rank_text: pbar leaked on CUDA OOM (executor was closed, pbar was not)
- rank_folder: same pbar leak on OOM

Also removed unused imports: math and Dict.

The try/except/raise pattern in extract_features now properly closes both
resources before re-raising. OOM handlers in rank_text/rank_folder
now also call pbar.close() before raising.

chore: version bump 0.0.0.3 → 0.0.1.0 077609b43e

fix: normalize text score difference to [-1,1] before cosine_to_score mapping 9ba0b25158

avg_pos - avg_neg ∈ [-2, 2] was passed directly to cosine_to_score which
maps [-1,1] → [0,1], producing out-of-range scores. Normalize by /2 first so
the full [-2,2] range maps correctly to [0,1].

perf: rolling window prefetch — 2 batches in-flight at all times f19eecbb35

Before: 1 future in-flight → GPU waits for next batch's load to finish
After: 2 futures queued → next batch is always already loaded when GPU finishes

Applies to extract_features, rank_text, and rank_folder.

fix: catch all exceptions in rank_text/rank_folder encode_images calls e16a97ee45

The narrow OOM-only catch leaked pbar and executor on any other
exception (RuntimeError, CUDA errors, open_clip internal errors).
Changed to bare `except Exception: ... raise` matching extract_features.

sadasd 345455a59c

Disable PIL DecompressionBombWarning 492f9cf7e5

Add --trim to rank and rank-text: keep top N%, move rest to ./delete 1c95306941

fdsfds e2a2bb5e77

find_images: skip already-renamed [score]_ files to prevent double-scanning f20b17abb2

Fixes re-run and edge cases where scored files could be counted alongside originals.

Revert find_images skip [score]_ prefix — wasn't the issue, reverting to keep original behavior 60ed69914d

DEBUG: find_images prints raw vs image count and extension breakdown 6c281e7ba2

DEBUG: track update calls vs scored count in rank/rank-text f174ddd086

FIX: rank_folder pbar.update on empty batch uses valid_batch_paths not full batch_paths 12ffe2f5f1

FIX: rank_folder duplicate batch enqueue — same fix as c-radio commit 2ccdb88 a0f45b091a

The enqueue logic used batch_end to compute next_batch_start, but batch_end
doesn't change between iterations when multiple entries with the same batch_end
exist in the queue. This caused the same next batch to be re-enqueued repeatedly.
Fix: track last_enqueued_end to only enqueue batches that haven't been added yet.

FIX rank_text: same duplicate batch enqueue fix as rank_folder dc3bdb1231

- Track last_enqueued_end to prevent re-enqueueing same batch
- Use batch_start + batch_size instead of batch_end for next start
- Fix pbar.update on empty batch to use valid_batch_paths not batch_paths

docs: add MERGE_DESIGN.md — plan to unify siglip + c-radio with --model flag d422276435

docs: add SIPHON_DESIGN.md — design for siphon top-N feature b89d45ff47

docs: add UNIFIED_RANK_DESIGN.md — unified rank, rich naming, siphon 7d258f22bc

feat: full implementation — unified rank, rich naming, siphon 3fbbd302a8

feat: partial rank unification — rank_folder now accepts positives/negatives, mode auto-detected 46681041b1

feat: unify rank_text into rank_folder — single scoring function with auto-detected mode ebb899be74

fix: remove duplicate @torch.no_grad decorator e6c46274da

fix: remove dead code branch, fix extract_features return, update docstring 20eeb259d7

fix: --name matches centroid filename with extension, not metadata 6a23c1fd95

fix: list command — simple filename listing, no metadata overhead 83e456eaa5

fix: list command — path check skipped for list/test commands ff30525762

info: print rank mode before scoring begins a66ea58bdd