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/training/gen_test_vectors.py | 74 +++++++++++++++---------------------- 1 file changed, 29 insertions(+), 45 deletions(-) (limited to 'cnn_v3/training/gen_test_vectors.py') diff --git a/cnn_v3/training/gen_test_vectors.py b/cnn_v3/training/gen_test_vectors.py index 640971c..2eb889c 100644 --- a/cnn_v3/training/gen_test_vectors.py +++ b/cnn_v3/training/gen_test_vectors.py @@ -23,7 +23,7 @@ DEC0_IN, DEC0_OUT = 8, 4 ENC0_WEIGHTS = ENC0_IN * ENC0_OUT * 9 + ENC0_OUT # 724 ENC1_WEIGHTS = ENC1_IN * ENC1_OUT * 9 + ENC1_OUT # 296 -BN_WEIGHTS = BN_IN * BN_OUT * 1 + BN_OUT # 72 +BN_WEIGHTS = BN_IN * BN_OUT * 9 + BN_OUT # 584 (3x3 dilation=2) DEC1_WEIGHTS = DEC1_IN * DEC1_OUT * 9 + DEC1_OUT # 580 DEC0_WEIGHTS = DEC0_IN * DEC0_OUT * 9 + DEC0_OUT # 292 @@ -32,30 +32,8 @@ ENC1_OFFSET = ENC0_OFFSET + ENC0_WEIGHTS BN_OFFSET = ENC1_OFFSET + ENC1_WEIGHTS DEC1_OFFSET = BN_OFFSET + BN_WEIGHTS DEC0_OFFSET = DEC1_OFFSET + DEC1_WEIGHTS -TOTAL_F16 = DEC0_OFFSET + DEC0_WEIGHTS # 1964 + 292 = 2256? let me check -# 724 + 296 + 72 + 580 + 292 = 1964 ... actually let me recount -# ENC0: 20*4*9 + 4 = 720+4 = 724 -# ENC1: 4*8*9 + 8 = 288+8 = 296 -# BN: 8*8*1 + 8 = 64+8 = 72 -# DEC1: 16*4*9 + 4 = 576+4 = 580 -# DEC0: 8*4*9 + 4 = 288+4 = 292 -# Total = 724+296+72+580+292 = 1964 ... but HOWTO.md says 2064. Let me recheck. -# DEC1: 16*4*9 = 576 ... but the shader says Conv(16->4) which is IN=16, OUT=4 -# weight idx: o * DEC1_IN * 9 + i * 9 + ki where o4) = OUT*IN*K^2 = 4*16*9 = 576 + bias 4 = 580. HOWTO says 576+4=580 OK. -# Total = 724+296+72+580+292 = let me sum: 724+296=1020, +72=1092, +580=1672, +292=1964. -# Hmm, HOWTO.md says 2064. Let me recheck HOWTO weight table: -# enc0: 20*4*9=720 +4 = 724 -# enc1: 4*8*9=288 +8 = 296 -# bottleneck: 8*8*1=64 +8 = 72 -# dec1: 16*4*9=576 +4 = 580 -# dec0: 8*4*9=288 +4 = 292 -# Total = 724+296+72+580+292 = 1964 -# The HOWTO says 2064 but I get 1964... 100 difference. Possible typo in doc. -# I'll use the correct value derived from the formulas: 1964. +TOTAL_F16 = DEC0_OFFSET + DEC0_WEIGHTS +# 724 + 296 + 584 + 580 + 292 = 2476 (BN is now 3x3 dilation=2, was 72) # --------------------------------------------------------------------------- # Helpers @@ -140,35 +118,41 @@ def enc1_forward(enc0, w, gamma_lo, gamma_hi, beta_lo, beta_hi): def bottleneck_forward(enc1, w): """ - AvgPool2x2(enc1, clamp-border) + Conv(8->8, 1x1) + ReLU + AvgPool2x2(enc1, clamp-border) + Conv(8->8, 3x3, dilation=2) + ReLU → rgba32uint (f16, quarter-res). No FiLM. enc1: (hH, hW, 8) f32 — half-res + Matches cnn_v3_bottleneck.wgsl exactly. """ hH, hW = enc1.shape[:2] qH, qW = hH // 2, hW // 2 wo = BN_OFFSET - # AvgPool2x2 with clamp (matches load_enc1_avg in WGSL) - avg = np.zeros((qH, qW, BN_IN), dtype=np.float32) - for qy in range(qH): - for qx in range(qW): - s = np.zeros(BN_IN, dtype=np.float32) - for dy in range(2): - for dx in range(2): - hy = min(qy * 2 + dy, hH - 1) - hx = min(qx * 2 + dx, hW - 1) - s += enc1[hy, hx, :] - avg[qy, qx, :] = s * 0.25 - - # 1x1 conv (no spatial loop, just channel dot-product) + def load_enc1_avg(qy, qx): + """Avg-pool 2x2 from enc1 at quarter-res coord. Zero for OOB (matches WGSL).""" + if qy < 0 or qx < 0 or qy >= qH or qx >= qW: + return np.zeros(BN_IN, dtype=np.float32) + s = np.zeros(BN_IN, dtype=np.float32) + for dy in range(2): + for dx in range(2): + hy = min(qy * 2 + dy, hH - 1) + hx = min(qx * 2 + dx, hW - 1) + s += enc1[hy, hx, :] + return s * 0.25 + + # 3x3 conv with dilation=2 in quarter-res space out = np.zeros((qH, qW, BN_OUT), dtype=np.float32) for o in range(BN_OUT): - bias = get_w(w, wo, BN_OUT * BN_IN + o) - s = np.full((qH, qW), bias, dtype=np.float32) - for i in range(BN_IN): - wv = get_w(w, wo, o * BN_IN + i) - s += wv * avg[:, :, i] - out[:, :, o] = np.maximum(0.0, s) + bias = get_w(w, wo, BN_OUT * BN_IN * 9 + o) + for qy in range(qH): + for qx in range(qW): + s = bias + for ky in range(-1, 2): + for kx in range(-1, 2): + feat = load_enc1_avg(qy + ky * 2, qx + kx * 2) # dilation=2 + ki = (ky + 1) * 3 + (kx + 1) + for i in range(BN_IN): + s += get_w(w, wo, o * BN_IN * 9 + i * 9 + ki) * feat[i] + out[qy, qx, o] = max(0.0, s) return np.float16(out).astype(np.float32) # pack2x16float boundary -- cgit v1.2.3