feature/c-radio #3

Merged

isaac merged 25 commits from feature/c-radio into main

2026-05-15 18:10:15 +00:00

isaac commented

2026-05-15 18:10:10 +00:00

Owner

No description provided.

isaac added 25 commits

2026-05-15 18:10:10 +00:00

feat: C-RADIOv4-H single-model scoring at 1024×1024 6c39df73b4

Replace DINOv2 + SigLIP2 two-model pipeline with nvidia/C-RADIOv4-H:
- CRadioExtractor: transformers.AutoModel + CLIPImageProcessor
- 1024-dim features, aspect-preserving resize (not center-crop)
- Streaming centroid: O(1) memory, accumulates running mean
- OOM retry at half batch (min batch 4)
- --resolution default 1024 (paper's optimal zero-shot resolution)
- --auto-curate --top N --dest PATH: copy top N to folder
- Removed ScoringHead (trained on pseudo-scores, no new signal)
- Removed train-head, fine-tune, test-frozen commands
- Removed DINO/SigLIP models, CombinedFeatureExtractor
- 10 tests passing

Net: -800 lines of code

feat: C-RADIOv4-H single-model scoring pipeline 5ec3c9e9a6

C-RADIOv4-H replaces DINOv2 + SigLIP2 dual-model approach.
- CRadioExtractor: nvidia/C-RADIOv4-H at 1024x1024, 1024-dim output
- Streaming centroid: O(1) memory, arbitrary dataset size
- OOM retry: halves batch until success, no silent drops
- centroid-only scoring (ScoringHead removed — no new signal)
- auto-curate: copy top-N scored images to dest folder
- --resolution flag: 224-2048, default 1024
- 3 auto-fixed review issues incorporated

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

fix: SyntaxWarning from backslash in docstring path 680c09809f

fix: CLIPImageProcessor uses shortest_edge not longest_edge 1877990a17

fix: CLIPImageProcessor padding=True for batch-compatible shapes 0a04b61093

fix: CLIPImageProcessor uses explicit height/width for uniform batch shape 957e6e043d

fix: per-image processing for aspect-ratio-preserving C-RADIO features ba0106703e

fix: snap cropped pixel_values to multiple of 16 for C-RADIO 3a870316b1

fix: handle C-RADIO non-tuple output and validate feature dim e04d9441b1

fix: extract correct 1280-dim summary from C-RADIO multi-token output fc59894e77

C-RADIOv4-H produces a 2560-dim summary (2 teachers × 1280-dim each).
We now slice out[0][:, :1280] to get a single teacher representation
for centroid scoring. Also updates FEATURE_DIM from 1024 to 1280.

fix: rank command uses centroid from extract (models/cradio_features.pt) 7dafec1351

fix: squeeze centroid dim to handle torch.load shape mismatch 8d2cea7063

fix: matmul centroid shape squeeze mismatch 5fbf8b1acb

perf: batch all images in single forward pass, pad to largest per batch e4d285c3a9

docs: fix stale 1024 references to 1280 in comments and docstrings 00ad934292

perf: batch processor call for uniform-size images, no per-image loop a50e7e9d36

chore: default batch_size to 10 9b274c6bde

feat: add progress indicator to extract_features 7325f66505

fix: revert extract_batch to per-image loop (extract doesn't need batching) 0fc2438e80

perf: batched extract with per-image processor then single forward pass 4659c854bc

fix: revert extract_batch to per-image loop (aspect ratio batching broken) 9aba4d572e

fix: rank_batch matmul shape handling for centroid dot product 6026c09418