CNN v2: Restore per-layer kernel sizes support

Training:
- train_cnn_v2.py: Accept --kernel-sizes as comma-separated list
- CNNv2 model: Per-layer kernel sizes (e.g., [1,3,5])
- Single value replicates across layers (e.g., "3" → [3,3,3])

Export:
- export_cnn_v2_weights.py: Backward compatible with old checkpoints
- Handles both kernel_size (old) and kernel_sizes (new) format

Documentation:
- CNN_V2.md: Updated code examples and config format
- HOWTO.md: Updated training examples to show comma-separated syntax

Binary format: Already supports per-layer kernel sizes (no changes)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

2 files changed, 9 insertions, 6 deletions

diff --git a/doc/CNN_V2.md b/doc/CNN_V2.md
index 4612d7a..6242747 100644
--- a/doc/CNN_V2.md
+++ b/doc/CNN_V2.md
@@ -214,12 +214,15 @@ def compute_static_features(rgb, depth):
 
 ```python
 class CNNv2(nn.Module):
-    def __init__(self, kernel_size=3, num_layers=3):
+    def __init__(self, kernel_sizes, num_layers=3):
         super().__init__()
+        if isinstance(kernel_sizes, int):
+            kernel_sizes = [kernel_sizes] * num_layers
+        self.kernel_sizes = kernel_sizes
         self.layers = nn.ModuleList()
 
         # All layers: 12D input (4 prev + 8 static) → 4D output
-        for i in range(num_layers):
+        for kernel_size in kernel_sizes:
             self.layers.append(
                 nn.Conv2d(12, 4, kernel_size=kernel_size,
                          padding=kernel_size//2, bias=False)
@@ -247,7 +250,7 @@ class CNNv2(nn.Module):
 
 ```python
 # Hyperparameters
-kernel_size = 3              # Uniform 3×3 kernels
+kernel_sizes = [3, 3, 3]     # Per-layer kernel sizes (e.g., [1,3,5])
 num_layers = 3               # Number of CNN layers
 learning_rate = 1e-3
 batch_size = 16
@@ -278,7 +281,7 @@ for epoch in range(epochs):
 torch.save({
     'state_dict': model.state_dict(),  # f32 weights
     'config': {
-        'kernel_size': 3,
+        'kernel_sizes': [3, 3, 3],  # Per-layer kernel sizes
         'num_layers': 3,
         'features': ['p0', 'p1', 'p2', 'p3', 'uv.x', 'uv.y', 'sin10_x', 'bias']
     },
diff --git a/doc/HOWTO.md b/doc/HOWTO.md
index e909a5d..9c67106 100644
--- a/doc/HOWTO.md
+++ b/doc/HOWTO.md
@@ -161,10 +161,10 @@ Config: 100 epochs, 3×3 kernels, 8→4→4 channels, patch-based (harris detect
   --input training/input/ --target training/target_2/ \
   --epochs 100 --batch-size 16 --checkpoint-every 5
 
-# Custom architecture
+# Custom architecture (per-layer kernel sizes)
 ./training/train_cnn_v2.py \
   --input training/input/ --target training/target_2/ \
-  --kernel-sizes 1 3 5 --channels 16 8 4 \
+  --kernel-sizes 1,3,5 \
   --epochs 5000 --batch-size 16
 ```