diff options
| author | skal <pascal.massimino@gmail.com> | 2026-03-25 10:05:42 +0100 |
|---|---|---|
| committer | skal <pascal.massimino@gmail.com> | 2026-03-25 10:05:42 +0100 |
| commit | ce6e5b99f26e4e7c69a3cacf360bd0d492de928c (patch) | |
| tree | a8d64b33a7ea1109b6b7e1043ced946cac416756 /cnn_v3/training/export_cnn_v3_weights.py | |
| parent | 8b4d7a49f038d7e849e6764dcc3abd1e1be01061 (diff) | |
feat(cnn_v3): 3×3 dilated bottleneck + Sobel loss + FiLM warmup + architecture PNG
- 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).
Diffstat (limited to 'cnn_v3/training/export_cnn_v3_weights.py')
| -rw-r--r-- | cnn_v3/training/export_cnn_v3_weights.py | 16 |
1 files changed, 8 insertions, 8 deletions
diff --git a/cnn_v3/training/export_cnn_v3_weights.py b/cnn_v3/training/export_cnn_v3_weights.py index edf76e2..78f5f25 100644 --- a/cnn_v3/training/export_cnn_v3_weights.py +++ b/cnn_v3/training/export_cnn_v3_weights.py @@ -15,8 +15,8 @@ Outputs <output_dir>/cnn_v3_weights.bin Conv+bias weights for all 5 passes, packed as f16-pairs-in-u32. Matches the format expected by CNNv3Effect::upload_weights(). - Layout: enc0 (724) | enc1 (296) | bottleneck (72) | dec1 (580) | dec0 (292) - = 1964 f16 values = 982 u32 = 3928 bytes. + Layout: enc0 (724) | enc1 (296) | bottleneck (584) | dec1 (580) | dec0 (292) + = 2476 f16 values = 1238 u32 = 4952 bytes. <output_dir>/cnn_v3_film_mlp.bin FiLM MLP weights as raw f32: L0_W (5×16) L0_b (16) L1_W (16×40) L1_b (40). @@ -48,13 +48,13 @@ from train_cnn_v3 import CNNv3 # cnn_v3/src/cnn_v3_effect.cc (kEnc0Weights, kEnc1Weights, …) # cnn_v3/training/gen_test_vectors.py (same constants) # --------------------------------------------------------------------------- -ENC0_WEIGHTS = 20 * 4 * 9 + 4 # Conv(20→4,3×3)+bias = 724 -ENC1_WEIGHTS = 4 * 8 * 9 + 8 # Conv(4→8,3×3)+bias = 296 -BN_WEIGHTS = 8 * 8 * 1 + 8 # Conv(8→8,1×1)+bias = 72 -DEC1_WEIGHTS = 16 * 4 * 9 + 4 # Conv(16→4,3×3)+bias = 580 -DEC0_WEIGHTS = 8 * 4 * 9 + 4 # Conv(8→4,3×3)+bias = 292 +ENC0_WEIGHTS = 20 * 4 * 9 + 4 # Conv(20→4,3×3)+bias = 724 +ENC1_WEIGHTS = 4 * 8 * 9 + 8 # Conv(4→8,3×3)+bias = 296 +BN_WEIGHTS = 8 * 8 * 9 + 8 # Conv(8→8,3×3,dil=2)+bias = 584 +DEC1_WEIGHTS = 16 * 4 * 9 + 4 # Conv(16→4,3×3)+bias = 580 +DEC0_WEIGHTS = 8 * 4 * 9 + 4 # Conv(8→4,3×3)+bias = 292 TOTAL_F16 = ENC0_WEIGHTS + ENC1_WEIGHTS + BN_WEIGHTS + DEC1_WEIGHTS + DEC0_WEIGHTS -# = 1964 +# = 2476 def pack_weights_u32(w_f16: np.ndarray) -> np.ndarray: |