summaryrefslogtreecommitdiff
path: root/cnn_v3/training
AgeCommit message (Collapse)Author
8 hoursfeat(cnn_v3): patch alignment search, resume, Ctrl-C saveskal
- --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.
8 hoursnormalize sample dimensionskal
9 hoursdocs(cnn_v3): add uv inline deps to train_cnn_v3.py + HOW_TO_CNN noteskal
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.
9 hoursperf(cnn_v3): cache dataset images at init to avoid per-patch disk I/Oskal
handoff(Gemini): CNNv3Dataset now loads all samples once in __init__ into self._cache; __getitem__ reads from cache instead of reloading PNGs each call. Eliminates N×patches_per_image file loads per epoch.
9 hoursfix(cnn_v3): correct weight budget in docstring (3.9→5.4 KB f16)skal
9 hoursfix(cnn_v3): resize target to albedo dims when sizes differskal
target.png can have a different resolution than albedo.png in simple samples; patch slicing into the smaller target produced 0×0 tensors, crashing torch.stack in the DataLoader collate. handoff(Gemini): target resized in _load_sample (LANCZOS) + note in HOW_TO_CNN §1c.
9 hoursdocs(cnn_v3): add full Old House example to HOW_TO_CNN §1bskal
handoff(Gemini): added render + batch-pack example commands at end of section 1b
10 hoursfix(cnn_v3): native OPEN_EXR_MULTILAYER + quiet render + flexible channel namesskal
blender_export.py: - Replace broken compositor FileOutput approach with native OPEN_EXR_MULTILAYER render output; all enabled passes included automatically, no socket wiring needed - Suppress Fra:/Mem: render spam via os.dup2 fd redirect; per-frame progress printed to stderr via render_post handler pack_blender_sample.py: - get_pass_r: try .R/.X/.Y/.Z/'' suffixes + aliases param for Depth→Z fallback - combined_rgba loaded once via ("Combined","Image") loop; shared by transp+target - Remove unused sys import HOW_TO_CNN.md: update channel table to native EXR naming (Depth.Z, IndexOB.X, Shadow.X), fix example command, note Shadow defaults to 255 when absent handoff(Gemini): blender pipeline now produces correct multilayer EXR with all G-buffer passes; pack script handles native channel naming
10 hoursfix(blender_export): version detection + Blender 5.x warning + cleanupskal
- Use bpy.app.version for version detection instead of attribute sniffing - Blender 5.0.x: warn that per-pass compositor routing is broken (Combined only); compositing_node_group path kept ready for when Blender fixes this upstream - Remove all DEBUG prints and failed use_nodes=True experiment - configure_scene() returns only discard_dir (compositor always configured) - Move _SOCKET_ALIASES to module level; simplify slots/None fallback handoff(Gemini): blender_export.py stable for Blender 4.5 LTS; Blender 5.x path is forward-compatible but produces Combined-only output until upstream fix.
11 hoursfix(cnn_v3): blender_export Blender 5 compositor activation + document ↵skal
RenderLayer sockets - Activate compositor in Blender 5.0+ by relying on compositing_node_group assignment (no use_nodes needed, avoids deprecation warning) - Document full CompositorNodeRLayers output socket list for Blender 5.0.1 - Clean up SOCKET_ALIASES to match confirmed socket names
11 hoursfeat(cnn_v3): blender_export print pack_blender_sample.py batch command ↵skal
after render
11 hoursfix(cnn_v3): blender_export fallback socket name aliases for Shadow etc.skal
11 hoursfix(cnn_v3): blender_export discard dir next to --output, not in /tmpskal
11 hoursfix(cnn_v3): blender_export.py Blender 5 File Output node slots + file_nameskal
- Prefer file_output_items over file_slots; use explicit is-None checks so empty collections do not fall through to the legacy attribute. - Clear out_node.file_name so multilayer EXR frames are named 0001.exr instead of file_name0001.exr. handoff(Gemini): blender_export.py now produces frames/0001.exr on Blender 5.0.1.
12 hoursdocs(cnn_v3): clarify --output is a base dir, not a frame_### patternskal
12 hoursfix(cnn_v3): blender_export.py Blender 5.x API compatibilityskal
- compositor: use compositing_node_group (Blender 5+) / node_tree (<=4.x) - file output: use file_output_items.new(type, name) (5+) / file_slots (older) - file output: use directory attr (5+) / base_path (older) - suppress default PNG output via mkdtemp + shutil.rmtree after render - link passes by name instead of positional index - add TODO for Shadow socket name variance across blend files - clean up: extract helpers, PASS_SOCKETS constant with socket types handoff(Gemini): blender_export.py now works on Blender 5.0.1
12 hoursfix(cnn_v3): blender_export --view-layer flag + fallback to layer[0]skal
Fixes KeyError when blend file uses a non-default view layer name. Adds --view-layer NAME arg; pass '?' to list available layers. Defaults to index 0 with a clear error if the name is not found. handoff(Gemini): blender_export.py view layer selection now robust
12 hoursfeat(cnn_v3): gen_sample tool + 7 simple training samplesskal
- pack_photo_sample.py: --target now required (no albedo fallback) - gen_sample.py: bash wrapper with positional args (input target output_dir) - input/photo7.jpg: copy of photo2 (second style target) - target_1: photo2_1_out→photo2_out, photo2_2_out→photo7_out - dataset/simple/sample_001..007: 7 packed photo/target pairs handoff(Gemini): training data ready; next step is train_cnn_v3.py run
12 hoursfeat(cnn_v3): gen_sample tool + 7 simple training samplesskal
- pack_photo_sample.py: --target now required (no albedo fallback) - gen_sample.py: bash wrapper with positional args (input target output_dir) - input/photo7.jpg: copy of photo2 (second style target) - target_1: photo2_1_out→photo2_out, photo2_2_out→photo7_out - dataset/simple/sample_001..007: 7 packed photo/target pairs handoff(Gemini): training data ready; next step is train_cnn_v3.py run
13 hoursfeat(cnn_v3): gen_sample tool + 7 simple training samplesskal
- pack_photo_sample.py: --target now required (no albedo fallback) - gen_sample: bash wrapper with positional args (input target output_dir) - input/photo7.jpg: copy of photo2 (second style target) - target_1: photo2_1_out→photo2_out, photo2_2_out→photo7_out - dataset/simple/sample_001..007: 7 packed photo/target pairs handoff(Gemini): training data ready; next step is train_cnn_v3.py run
30 hoursrefactor(cnn_v3): code review — comments, simplifications, test fixskal
C++: - cnn_v3_effect.cc: fix declare_nodes comment (output node declared by caller) - cnn_v3_effect.cc: add TODO(phase-7) marker for FiLM MLP replacement WGSL: - cnn_v3_bottleneck.wgsl: consolidate _pad fields onto one line, explain why array<u32,3> is invalid in uniform address space - cnn_v3_enc0.wgsl: fix "12xu8" → "12ch u8norm" in header comment - cnn_v3_dec0.wgsl: clarify parity note (sigmoid after FiLM+ReLU, not raw conv) - cnn_v3_common.wgsl: clarify unpack_8ch pack layout (low/high 16 bits) Python: - cnn_v3_utils.py: replace PIL-based _upsample_nearest (uint8 round-trip) with pure numpy index arithmetic; rename _resize_rgb → _resize_img (handles any channel count); add comment on normal zero-pad workaround - export_cnn_v3_weights.py: add cross-ref to cnn_v3_effect.cc constants; clarify weight count comments with Conv notation Test: - test_cnn_v3_parity.cc: enc0/dec1 layer failures now return 0 (were print-only) handoff(Gemini): CNN v3 review complete, 36/36 tests passing.
34 hoursfeat(cnn_v3): export script + HOW_TO_CNN.md playbookskal
- export_cnn_v3_weights.py: .pth → cnn_v3_weights.bin (f16 packed u32) + cnn_v3_film_mlp.bin (f32) - HOW_TO_CNN.md: full pipeline playbook (data collection, training, export, C++ wiring, parity, HTML tool) - TODO.md: mark export script done Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
34 hoursfeat(cnn_v3): Phase 6 — training script (train_cnn_v3.py + cnn_v3_utils.py)skal
- 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>
34 hoursfeat(cnn_v3): Phase 5 complete — parity validation passing (36/36 tests)skal
- Add test_cnn_v3_parity.cc: zero_weights + random_weights tests - Add gen_test_vectors.py: PyTorch reference implementation for enc0/enc1/bn/dec1/dec0 - Add test_vectors.h: generated C header with enc0, dec1, output expected values - Fix declare_nodes(): intermediate textures at fractional resolutions (W/2, W/4) using new NodeRegistry::default_width()/default_height() getters - Add layer-by-layer readback (enc0, dec1) for regression coverage - Final parity: enc0 max_err=1.95e-3, dec1 max_err=1.95e-3, out max_err=4.88e-4 handoff(Claude): CNN v3 parity done. Next: train_cnn_v3.py (FiLM MLP training).
2 daysfeat(cnn_v3): G-buffer phase 1 + training infrastructureskal
G-buffer (Phase 1): - Add NodeTypes GBUF_ALBEDO/DEPTH32/R8/RGBA32UINT to NodeRegistry - GBufferEffect: MRT raster pass (albedo+normal_mat+depth) + pack compute - Shaders: gbuf_raster.wgsl (MRT), gbuf_pack.wgsl (feature packing, 32B/px) - Shadow/SDF passes stubbed (placeholder textures), CMake integration deferred Training infrastructure (Phase 2): - blender_export.py: headless EXR export with all G-buffer render passes - pack_blender_sample.py: EXR → per-channel PNGs (oct-normals, 1/z depth) - pack_photo_sample.py: photo → zero-filled G-buffer sample layout handoff(Gemini): G-buffer phases 3-5 remain (U-Net shaders, CNNv3Effect, parity)
2026-03-05add training photosskal
2026-02-27remove old files, add new training setskal
2026-02-15feat(cnn): add CNN v3 directory structure with training dataskal
Initialize CNN v3 subdirectory with training pipeline layout: - docs/, scripts/, shaders/, src/, tools/, weights/ for organization - training/input/ with sample images - training/target_1/, target_2/ for multi-style training - README.md documenting structure Training images tracked in repo for easy collaboration. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
generated by cgit v1.2.3 (git 2.51.0) at 2026-03-22 19:03:36 +0000