| Age | Commit message (Collapse) | Author |
|---|
|
architecture PNG
- Replace 1×1 pointwise bottleneck with Conv(8→8, 3×3, dilation=2):
effective RF grows from ~13px to ~29px at ¼res (~+1 KB weights)
- Add Sobel edge loss in training (--edge-loss-weight, default 0.1)
- Add FiLM 2-phase training: freeze MLP for warmup epochs then
unfreeze at lr×0.1 (--film-warmup-epochs, default 50)
- Update weight layout: BN 72→584 f16, total 1964→2476 f16 (4952 B)
- Cascade offsets in C++ effect, JS tool, export/gen_test_vectors scripts
- Regenerate test_vectors.h (1238 u32); parity max_err=9.77e-04
- Generate dark-theme U-Net+FiLM architecture PNG (gen_architecture_png.py)
- Replace ASCII art in CNN_V3.md and HOW_TO_CNN.md with PNG embed
handoff(Gemini): bottleneck dilation + Sobel loss + FiLM warmup landed.
Next: run first real training pass (see cnn_v3/docs/HOWTO.md §3).
|
|
- train_cnn_v3.py: --single-sample <dir> implies --full-image + --batch-size 1
- cnn_v3_utils.py: CNNv3Dataset accepts single_sample= kwarg (explicit override)
- HOWTO.md: document --single-sample workflow, fix pack_photo_sample.py usage (--target required)
- HOW_TO_CNN.md: fix GBufferEffect seq input (prev_cnn→source), fix binary name (demo→demo64k), add --resume to flag table, remove stale "pack without target" block
handoff(Gemini): --single-sample <dir> added to train_cnn_v3.py; docs audited and corrected
|
|
- --patch-search-window N: at dataset init, find per-patch (dx,dy) in
[-N,N]² that minimises grayscale MSE between source albedo and target;
result cached so __getitem__ pays only a list-lookup per sample.
- --resume [CKPT]: restore model + Adam state from a checkpoint; omit
path to auto-select the latest in --checkpoint-dir.
- Ctrl-C (SIGINT) finishes the current batch, then saves a checkpoint
before exiting; finally-block guarded so no spurious epoch-0 save.
- Review: remove unused sd variable, lift patch_idx out of duplicate
computation, move _LUMA to Constants block, update module docstring.
handoff(Gemini): cnn_v3/training updated — no C++ or test changes.
|
|
handoff(Gemini): train_cnn_v3.py now has uv script metadata block
(torch, torchvision, numpy, pillow, opencv-python). HOW_TO_CNN §2
Prerequisites updated with uv quick-start alternative.
|
|
|
|
- train_cnn_v3.py: CNNv3 U-Net+FiLM model, training loop, CLI
- cnn_v3_utils.py: image I/O, pyrdown, depth_gradient, assemble_features,
apply_channel_dropout, detect_salient_points, CNNv3Dataset
- Patch-based training (default 64×64) with salient-point extraction
(harris/shi-tomasi/fast/gradient/random detectors, pre-cached at init)
- Channel dropout for geometric/context/temporal channels
- Random FiLM conditioning per sample for joint MLP+U-Net training
- docs: HOWTO.md §3 updated with commands and flag reference
- TODO.md: Phase 6 marked done, export script noted as next step
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
|
