diff options
| author | skal <pascal.massimino@gmail.com> | 2026-02-11 09:27:06 +0100 |
|---|---|---|
| committer | skal <pascal.massimino@gmail.com> | 2026-02-11 09:27:06 +0100 |
| commit | 5cc6da3831d5bce35af353c14f15c30dbc66b081 (patch) | |
| tree | 800526fe4c3a937121d7f1e87b85ff89cc3c7fcb /doc | |
| parent | 66a489f64209925ec9615c9f6c4907e4e3caf9e2 (diff) | |
fix: CNN training/inference to match WGSL sliding window
Training now computes loss only on center pixels (excludes conv padding
borders). Inference changed from tiling to full-image sliding window.
Both match cnn_layer.wgsl: each pixel processed from NxN neighborhood.
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
Diffstat (limited to 'doc')
| -rw-r--r-- | doc/HOWTO.md | 15 |
1 files changed, 11 insertions, 4 deletions
diff --git a/doc/HOWTO.md b/doc/HOWTO.md index db324ec..7b0daa0 100644 --- a/doc/HOWTO.md +++ b/doc/HOWTO.md @@ -89,7 +89,7 @@ make run_util_tests # Utility tests ## Training ### Patch-Based (Recommended) -Extracts patches at salient points, preserves natural pixel scale: +Extracts patches at salient points, trains on center pixels only (matches WGSL sliding window): ```bash # Train with 32×32 patches at detected corners/edges ./training/train_cnn.py \ @@ -99,10 +99,15 @@ Extracts patches at salient points, preserves natural pixel scale: --checkpoint-every 1000 ``` +**Training behavior:** +- Loss computed only on center pixels (excludes conv padding borders) +- For 3-layer network: excludes 3px border on each side +- Matches GPU shader sliding-window paradigm + **Detectors:** `harris` (default), `fast`, `shi-tomasi`, `gradient` -### Full-Image (Legacy) -Resizes to 256×256 (distorts scale): +### Full-Image +Processes entire image with sliding window (matches WGSL): ```bash ./training/train_cnn.py \ --input training/input/ --target training/output/ \ @@ -115,12 +120,14 @@ Resizes to 256×256 (distorts scale): # Generate shaders from checkpoint ./training/train_cnn.py --export-only checkpoints/checkpoint_epoch_5000.pth -# Generate ground truth for comparison +# Generate ground truth (sliding window, no tiling) ./training/train_cnn.py --infer input.png \ --export-only checkpoints/checkpoint_epoch_5000.pth \ --output ground_truth.png ``` +**Inference:** Processes full image with sliding window (each pixel from NxN neighborhood). No tiling artifacts. + **Kernel sizes:** 3×3 (36 weights), 5×5 (100 weights), 7×7 (196 weights) --- |