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/train_cnn_v3.py | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) (limited to 'cnn_v3/training/train_cnn_v3.py') diff --git a/cnn_v3/training/train_cnn_v3.py b/cnn_v3/training/train_cnn_v3.py index e48f684..fa0d2e2 100644 --- a/cnn_v3/training/train_cnn_v3.py +++ b/cnn_v3/training/train_cnn_v3.py @@ -10,8 +10,7 @@ Architecture (enc_channels=[8,16]): enc1 Conv(8→16, 3×3) + FiLM + ReLU + pool2 H/2×W/2 2× rgba32uint (16ch split) bottleneck Conv(16→16, 3×3, dilation=2) + ReLU H/4×W/4 2× rgba32uint (16ch split) dec1 upsample×2 + cat(enc1) Conv(32→8) + FiLM H/2×W/2 rgba32uint (8ch) - dec0 upsample×2 + cat(enc0) Conv(16→4) + FiLM H×W rgba16float (4ch) - output sigmoid → RGBA + dec0 upsample×2 + cat(enc0) Conv(16→4) + FiLM + sigmoid H×W rgba16float (4ch) FiLM MLP: Linear(5→16) → ReLU → Linear(16→72) 72 = 2 × (γ+β) for enc0(8) enc1(16) dec1(8) dec0(4) @@ -93,9 +92,9 @@ class CNNv3(nn.Module): torch.cat([F.interpolate(x, scale_factor=2, mode='nearest'), skip1], dim=1) ), gd1, bd1)) - x = F.relu(film_apply(self.dec0( + x = film_apply(self.dec0( torch.cat([F.interpolate(x, scale_factor=2, mode='nearest'), skip0], dim=1) - ), gd0, bd0)) + ), gd0, bd0) return torch.sigmoid(x) -- cgit v1.2.3