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/docs/CNN_V3.md | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) (limited to 'cnn_v3/docs/CNN_V3.md') diff --git a/cnn_v3/docs/CNN_V3.md b/cnn_v3/docs/CNN_V3.md index a197a1d..081adf8 100644 --- a/cnn_v3/docs/CNN_V3.md +++ b/cnn_v3/docs/CNN_V3.md @@ -19,7 +19,7 @@ CNN v3 is a next-generation post-processing effect using: - Training from both Blender renders and real photos - Strict test framework: per-pixel bit-exact validation across all implementations -**Status:** Phases 1–7 complete. Architecture upgraded to enc_channels=[8,16] for improved capacity. Parity test and runtime updated. Next: training pass. +**Status:** Phases 1–9 complete. Architecture upgraded to enc_channels=[8,16]. Two training bugs fixed (dec0 ReLU removed; FiLM MLP loaded at runtime). Parity validated. Next: retrain from scratch with more data. --- @@ -34,9 +34,13 @@ FiLM is applied **inside each encoder/decoder block**, after each convolution. ### U-Net Block (per level) ``` -input → Conv 3×3 → BN (or none) → FiLM(γ,β) → ReLU → output +enc0/enc1/dec1: input → Conv 3×3 → FiLM(γ,β) → ReLU → output +dec0 (final): input → Conv 3×3 → FiLM(γ,β) → Sigmoid → output ``` +The final decoder layer uses sigmoid directly — **no ReLU** — so the network +can output the full [0,1] range. ReLU before sigmoid would clamp to [0.5,1.0]. + FiLM at level `l`: ``` FiLM(x, γ_l, β_l) = γ_l ⊙ x + β_l (per-channel affine) -- cgit v1.2.3