diff options
Diffstat (limited to 'doc/CNN_V2.md')
| -rw-r--r-- | doc/CNN_V2.md | 43 |
1 files changed, 40 insertions, 3 deletions
diff --git a/doc/CNN_V2.md b/doc/CNN_V2.md index 78854ce..abef606 100644 --- a/doc/CNN_V2.md +++ b/doc/CNN_V2.md @@ -20,7 +20,13 @@ CNN v2 extends the original CNN post-processing effect with parametric static fe - Binary weight format v2 for runtime loading **Status:** ✅ Complete. Training pipeline functional, validation tools ready, mip-level support integrated. -**TODO:** 8-bit quantization with QAT for 2× size reduction (~1.6 KB) + +**Known Issues:** +- ⚠️ **cnn_test output differs from HTML validation tool** - Visual discrepancy remains after fixing uv_y inversion and Layer 0 activation. Root cause under investigation. Both tools should produce identical output given same weights/input. + +**TODO:** +- 8-bit quantization with QAT for 2× size reduction (~1.6 KB) +- Debug cnn_test vs HTML tool output difference --- @@ -130,6 +136,27 @@ let bias = 1.0; // Learned bias per output channel // Packed storage: [p0, p1, p2, p3, uv.x, uv.y, sin(20*uv.y), 1.0] ``` +### Input Channel Mapping + +**Weight tensor layout (12 input channels per layer):** + +| Input Channel | Feature | Description | +|--------------|---------|-------------| +| 0-3 | Previous layer output | 4D RGBA from prior CNN layer (or input RGBD for Layer 0) | +| 4-11 | Static features | 8D: p0, p1, p2, p3, uv_x, uv_y, sin20_y, bias | + +**Static feature channel details:** +- Channel 4 → p0 (RGB.r from mip level) +- Channel 5 → p1 (RGB.g from mip level) +- Channel 6 → p2 (RGB.b from mip level) +- Channel 7 → p3 (depth or RGB channel from mip level) +- Channel 8 → p4 (uv_x: normalized horizontal position) +- Channel 9 → p5 (uv_y: normalized vertical position) +- Channel 10 → p6 (sin(20*uv_y): periodic encoding) +- Channel 11 → p7 (bias: constant 1.0) + +**Note:** When generating identity weights, p4-p7 correspond to input channels 8-11, not 4-7. + ### Feature Rationale | Feature | Dimension | Purpose | Priority | @@ -326,12 +353,13 @@ class CNNv2(nn.Module): kernel_sizes = [3, 3, 3] # Per-layer kernel sizes (e.g., [1,3,5]) num_layers = 3 # Number of CNN layers mip_level = 0 # Mip level for p0-p3: 0=orig, 1=half, 2=quarter, 3=eighth +grayscale_loss = False # Compute loss on grayscale (Y) instead of RGBA learning_rate = 1e-3 batch_size = 16 epochs = 5000 # Dataset: Input RGB, Target RGBA (preserves alpha channel from image) -# Model outputs RGBA, loss compares all 4 channels +# Model outputs RGBA, loss compares all 4 channels (or grayscale if --grayscale-loss) # Training loop (standard PyTorch f32) for epoch in range(epochs): @@ -344,7 +372,15 @@ for epoch in range(epochs): # Forward pass output = model(input_rgbd, static_feat) - loss = criterion(output, target_batch) + + # Loss computation (grayscale or RGBA) + if grayscale_loss: + # Convert RGBA to grayscale: Y = 0.299*R + 0.587*G + 0.114*B + output_gray = 0.299 * output[:, 0:1] + 0.587 * output[:, 1:2] + 0.114 * output[:, 2:3] + target_gray = 0.299 * target[:, 0:1] + 0.587 * target[:, 1:2] + 0.114 * target[:, 2:3] + loss = criterion(output_gray, target_gray) + else: + loss = criterion(output, target_batch) # Backward pass optimizer.zero_grad() @@ -361,6 +397,7 @@ torch.save({ 'kernel_sizes': [3, 3, 3], # Per-layer kernel sizes 'num_layers': 3, 'mip_level': 0, # Mip level used for p0-p3 + 'grayscale_loss': False, # Whether grayscale loss was used 'features': ['p0', 'p1', 'p2', 'p3', 'uv.x', 'uv.y', 'sin10_x', 'bias'] }, 'epoch': epoch, |