From ce6e5b99f26e4e7c69a3cacf360bd0d492de928c Mon Sep 17 00:00:00 2001 From: skal Date: Wed, 25 Mar 2026 10:05:42 +0100 Subject: feat(cnn_v3): 3×3 dilated bottleneck + Sobel loss + FiLM warmup + architecture PNG MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - 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). --- cnn_v3/src/cnn_v3_effect.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'cnn_v3/src') diff --git a/cnn_v3/src/cnn_v3_effect.cc b/cnn_v3/src/cnn_v3_effect.cc index bfbb17b..1391eba 100644 --- a/cnn_v3/src/cnn_v3_effect.cc +++ b/cnn_v3/src/cnn_v3_effect.cc @@ -25,7 +25,7 @@ // static const uint32_t kEnc0Weights = 20 * 4 * 9 + 4; // Conv(20→4,3×3)+bias static const uint32_t kEnc1Weights = 4 * 8 * 9 + 8; // Conv(4→8,3×3)+bias -static const uint32_t kBnWeights = 8 * 8 * 1 + 8; // Conv(8→8,1×1)+bias +static const uint32_t kBnWeights = 8 * 8 * 9 + 8; // Conv(8→8,3×3,dilation=2)+bias static const uint32_t kDec1Weights = 16 * 4 * 9 + 4; // Conv(16→4,3×3)+bias static const uint32_t kDec0Weights = 8 * 4 * 9 + 4; // Conv(8→4,3×3)+bias -- cgit v1.2.3