diff options
Diffstat (limited to 'cnn_v3/docs/HOW_TO_CNN.md')
| -rw-r--r-- | cnn_v3/docs/HOW_TO_CNN.md | 32 |
1 files changed, 16 insertions, 16 deletions
diff --git a/cnn_v3/docs/HOW_TO_CNN.md b/cnn_v3/docs/HOW_TO_CNN.md index 09db97c..11ed260 100644 --- a/cnn_v3/docs/HOW_TO_CNN.md +++ b/cnn_v3/docs/HOW_TO_CNN.md @@ -358,7 +358,7 @@ uv run train_cnn_v3.py \ The model prints its parameter count: ``` -Model: enc=[4, 8] film_cond_dim=5 params=3252 (~6.4 KB f16) +Model: enc=[8, 16] film_cond_dim=5 params=9148 (~17.9 KB f16) ``` If `params` is much higher, `--enc-channels` was changed; update C++ constants accordingly. @@ -492,12 +492,12 @@ WEIGHTS_CNN_V3_FILM_MLP, BINARY, weights/cnn_v3_film_mlp.bin, "CNN v3 FiLM MLP w | Layer | f16 count | Bytes | |-------|-----------|-------| -| enc0 Conv(20→4,3×3)+bias | 724 | — | -| enc1 Conv(4→8,3×3)+bias | 296 | — | -| bottleneck Conv(8→8,3×3,dil=2)+bias | 584 | — | -| dec1 Conv(16→4,3×3)+bias | 580 | — | -| dec0 Conv(8→4,3×3)+bias | 292 | — | -| **Total** | **2476 f16** | **4952 bytes** | +| enc0 Conv(20→8,3×3)+bias | 1448 | — | +| enc1 Conv(8→16,3×3)+bias | 1168 | — | +| bottleneck Conv(16→16,3×3,dil=2)+bias | 2320 | — | +| dec1 Conv(32→8,3×3)+bias | 2312 | — | +| dec0 Conv(16→4,3×3)+bias | 580 | — | +| **Total** | **7828 f16** | **15656 bytes** | **`cnn_v3_film_mlp.bin`** — FiLM MLP weights as raw f32, row-major: @@ -505,9 +505,9 @@ WEIGHTS_CNN_V3_FILM_MLP, BINARY, weights/cnn_v3_film_mlp.bin, "CNN v3 FiLM MLP w |-------|-------|-----------| | L0 weight | (16, 5) | 80 | | L0 bias | (16,) | 16 | -| L1 weight | (40, 16) | 640 | -| L1 bias | (40,) | 40 | -| **Total** | | **776 f32 = 3104 bytes** | +| L1 weight | (72, 16) | 1152 | +| L1 bias | (72,) | 72 | +| **Total** | | **1320 f32 = 5280 bytes** | The FiLM MLP is for CPU-side inference (future — see §4d). The U-Net weights in `cnn_v3_weights.bin` are what you need immediately. @@ -524,16 +524,16 @@ The export script produces this layout: `u32 = u16[0::2] | (u16[1::2] << 16)`. ``` Checkpoint: epoch=200 loss=0.012345 - enc_channels=[4, 8] film_cond_dim=5 + enc_channels=[8, 16] film_cond_dim=5 cnn_v3_weights.bin - 2476 f16 values → 1238 u32 → 4952 bytes - Upload via CNNv3Effect::upload_weights(queue, data, 4952) + 7828 f16 values → 3914 u32 → 15656 bytes + Upload via CNNv3Effect::upload_weights(queue, data, 15656) cnn_v3_film_mlp.bin L0: weight (16, 5) + bias (16,) - L1: weight (40, 16) + bias (40,) - 776 f32 values → 3104 bytes + L1: weight (72, 16) + bias (72,) + 1320 f32 values → 5280 bytes ``` ### Pitfalls @@ -542,7 +542,7 @@ cnn_v3_film_mlp.bin assertion in the export script fires. The C++ weight-offset constants (`kEnc0Weights` etc.) in `cnn_v3_effect.cc` must also be updated to match. - **Old checkpoint missing `config`:** if `config` key is absent (checkpoint from a very early - version), the script defaults to `enc_channels=[4,8], film_cond_dim=5`. + version), the script defaults to `enc_channels=[8,16], film_cond_dim=5`. - **`weights_only=True`:** requires PyTorch ≥ 2.0. If you get a warning, upgrade torch. --- |