From fb13e67acbc7d7dd2974a456fcb134966c47cee0 Mon Sep 17 00:00:00 2001 From: skal Date: Fri, 27 Mar 2026 07:59:00 +0100 Subject: fix(cnn_v3): remove dec0 ReLU, load FiLM MLP at runtime MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Two bugs blocking training convergence: 1. dec0 ReLU before sigmoid constrained output to [0.5,1.0] — network could never produce dark pixels. Removed F.relu in train_cnn_v3.py and max(0,…) in cnn_v3_dec0.wgsl. Test vectors regenerated. 2. set_film_params() used hardcoded heuristics instead of the trained MLP. Added CNNv3FilmMlp struct + load_film_mlp() to cnn_v3_effect.h/.cc. MLP auto-loaded from ASSET_WEIGHTS_CNN_V3_FILM_MLP at construction; Linear(5→16)→ReLU→Linear(16→72) runs CPU-side each frame. 36/36 tests pass. Parity max_err=4.88e-4 unchanged. handoff(Gemini): retrain from scratch — needs ≥50 samples (currently 11). See cnn_v3/docs/HOWTO.md §2-3. --- cnn_v3/training/gen_test_vectors.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 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 3f81247..96b175a 100644 --- a/cnn_v3/training/gen_test_vectors.py +++ b/cnn_v3/training/gen_test_vectors.py @@ -193,8 +193,8 @@ def dec1_forward(bn, enc1, w, gamma, beta): def dec0_forward(dec1, enc0, w, gamma, beta): """ - NearestUp2x(dec1) + cat(enc0_skip) → Conv(16->4, 3x3, zero-pad) + FiLM + ReLU + sigmoid - → rgba16float (full-res, final output). + NearestUp2x(dec1) + cat(enc0_skip) → Conv(16->4, 3x3, zero-pad) + FiLM + sigmoid + → rgba16float (full-res, final output). No ReLU before sigmoid. dec1: (hH, hW, 8) f32 — half-res enc0: (H, W, 8) f32 — full-res enc0 skip """ @@ -219,8 +219,8 @@ def dec0_forward(dec1, enc0, w, gamma, beta): for kx in range(3): wv = get_w(w, wo, o * DEC0_IN * 9 + i * 9 + ky * 3 + kx) s += wv * fp[ky:ky+H, kx:kx+W, i] - # FiLM + ReLU + sigmoid (matches WGSL dec0 shader) - v = np.maximum(0.0, gamma[o] * s + beta[o]) + # FiLM + sigmoid (matches WGSL dec0 shader — no ReLU before sigmoid) + v = gamma[o] * s + beta[o] out[:, :, o] = 1.0 / (1.0 + np.exp(-v.astype(np.float64))).astype(np.float32) return np.float16(out).astype(np.float32) # rgba16float boundary -- cgit v1.2.3