# CNN v2: Parametric Static Features

**Technical Design Document**

---

## Overview

CNN v2 extends the original CNN post-processing effect with parametric static features, enabling richer spatial and frequency-domain inputs for improved visual quality.

**Key improvements over v1:**
- 7D static feature input (vs 4D RGB)
- Multi-frequency position encoding (NeRF-style)
- Per-layer configurable kernel sizes (1×1, 3×3, 5×5)
- Variable channel counts per layer
- Float16 weight storage (~3.2 KB for 3-layer model)
- Bias integrated as static feature dimension
- Storage buffer architecture (dynamic layer count)
- Binary weight format for runtime loading

**Status:** ✅ Complete. Training pipeline functional, validation tools ready.
**TODO:** 8-bit quantization with QAT for 2× size reduction (~1.6 KB)

---

## Architecture

### Pipeline Overview

```
Input RGBD → Static Features Compute → CNN Layers → Output RGBA
             └─ computed once/frame ─┘  └─ multi-pass ─┘
```

**Static Features Texture:**
- Name: `static_features`
- Format: `texture_storage_2d<rgba32uint, write>` (4×u32)
- Data: 8 float16 values packed via `pack2x16float()`
- Computed once per frame, read by all CNN layers
- Lifetime: Entire frame (all CNN layer passes)

**CNN Layers:**
- Input Layer: 7D static features → C₀ channels
- Inner Layers: (7D + Cᵢ₋₁) → Cᵢ channels
- Output Layer: (7D + Cₙ) → 4D RGBA
- Storage: `texture_storage_2d<rgba32uint>` (8×f16 per texel recommended)

---

## Static Features (7D + 1 bias)

### Feature Layout

**8 float16 values per pixel:**

```wgsl
// Slot 0-3: RGBD (core pixel data)
let r = rgba.r;          // Red channel
let g = rgba.g;          // Green channel
let b = rgba.b;          // Blue channel
let d = depth;           // Depth value

// Slot 4-5: UV coordinates (normalized screen space)
let uv_x = coord.x / resolution.x;  // Horizontal position [0,1]
let uv_y = coord.y / resolution.y;  // Vertical position [0,1]

// Slot 6: Multi-frequency position encoding
let sin10_x = sin(10.0 * uv_x);     // Periodic feature (frequency=10)

// Slot 7: Bias dimension (always 1.0)
let bias = 1.0;                     // Learned bias per output channel

// Packed storage: [R, G, B, D, uv.x, uv.y, sin(10*uv.x), 1.0]
```

### Feature Rationale

| Feature | Dimension | Purpose | Priority |
|---------|-----------|---------|----------|
| RGBD | 4D | Core pixel information | Essential |
| UV coords | 2D | Spatial position awareness | Essential |
| sin(10\*uv.x) | 1D | Periodic position encoding | Medium |
| Bias | 1D | Learned bias (standard NN) | Essential |

**Why bias as static feature:**
- Simpler shader code (single weight array)
- Standard NN formulation: y = Wx (x includes bias term)
- Saves 56-112 bytes (no separate bias buffer)
- 7 features sufficient for initial implementation

### Future Feature Extensions

**Option: Replace sin(10\*uv.x) with:**
- `sin(20*uv.x)` - Higher frequency encoding
- `gray_mip1` - Multi-scale luminance
- `dx`, `dy` - Sobel gradients
- `variance` - Local texture measure
- `laplacian` - Edge detection

**Option: uint8 packing (16+ features):**
```wgsl
// texture_storage_2d<rgba8unorm> stores 16 uint8 values
// Trade precision for feature count
// [R, G, B, D, uv.x, uv.y, sin10.x, sin10.y,
//  sin20.x, sin20.y, dx, dy, gray_mip1, gray_mip2, var, bias]
```
Requires quantization-aware training.

---

## Layer Structure

### Example 3-Layer Network

```
Input:  7D static → 16 channels (1×1 kernel, pointwise)
Layer1: (7+16)D → 8 channels (3×3 kernel, spatial)
Layer2: (7+8)D → 4 channels (5×5 kernel, large receptive field)
```

### Weight Calculations

**Per-layer weights:**
```
Input:  7 × 1 × 1 × 16 = 112 weights
Layer1: (7+16) × 3 × 3 × 8 = 1656 weights
Layer2: (7+8) × 5 × 5 × 4 = 1500 weights
Total: 3268 weights
```

**Storage sizes:**
- f32: 3268 × 4 = 13,072 bytes (~12.8 KB)
- f16: 3268 × 2 = 6,536 bytes (~6.4 KB) ✓ **recommended**

**Comparison to v1:**
- v1: ~800 weights (3.2 KB f32)
- v2: ~3268 weights (6.4 KB f16)
- **Growth: 2× size for parametric features**

### Kernel Size Guidelines

**1×1 kernel (pointwise):**
- No spatial context, channel mixing only
- Weights: `(7 + C_in) × C_out`
- Use for: Input layer, bottleneck layers

**3×3 kernel (standard conv):**
- Local spatial context
- Weights: `(7 + C_in) × 9 × C_out`
- Use for: Most inner layers

**5×5 kernel (large receptive field):**
- Wide spatial context
- Weights: `(7 + C_in) × 25 × C_out`
- Use for: Output layer, detail enhancement

### Channel Storage (8×f16 per texel)

```wgsl
@group(0) @binding(1) var layer_input: texture_2d<u32>;

fn unpack_channels(coord: vec2<i32>) -> array<f32, 8> {
  let packed = textureLoad(layer_input, coord, 0);
  return array(
    unpack2x16float(packed.x).x, unpack2x16float(packed.x).y,
    unpack2x16float(packed.y).x, unpack2x16float(packed.y).y,
    unpack2x16float(packed.z).x, unpack2x16float(packed.z).y,
    unpack2x16float(packed.w).x, unpack2x16float(packed.w).y
  );
}

fn pack_channels(values: array<f32, 8>) -> vec4<u32> {
  return vec4(
    pack2x16float(vec2(values[0], values[1])),
    pack2x16float(vec2(values[2], values[3])),
    pack2x16float(vec2(values[4], values[5])),
    pack2x16float(vec2(values[6], values[7]))
  );
}
```

---

## Training Workflow

### Script: `training/train_cnn_v2.py`

**Static Feature Extraction:**

```python
def compute_static_features(rgb, depth):
    """Generate 7D static features + bias dimension."""
    h, w = rgb.shape[:2]

    # RGBD channels
    r, g, b = rgb[..., 0], rgb[..., 1], rgb[..., 2]

    # UV coordinates (normalized)
    uv_x = np.linspace(0, 1, w)[None, :].repeat(h, axis=0)
    uv_y = np.linspace(0, 1, h)[:, None].repeat(w, axis=1)

    # Multi-frequency position encoding
    sin10_x = np.sin(10.0 * uv_x)

    # Bias dimension (always 1.0)
    bias = np.ones_like(r)

    # Stack: [R, G, B, D, uv.x, uv.y, sin10_x, bias]
    return np.stack([r, g, b, depth, uv_x, uv_y, sin10_x, bias], axis=-1)
```

**Network Definition:**

```python
class CNNv2(nn.Module):
    def __init__(self, kernels=[1,3,5], channels=[16,8,4]):
        super().__init__()

        # Input layer: 8D (7 features + bias) → channels[0]
        self.layer0 = nn.Conv2d(8, channels[0], kernel_size=kernels[0],
                                padding=kernels[0]//2, bias=False)

        # Inner layers: (7 features + bias + C_prev) → C_next
        in_ch_1 = 8 + channels[0]  # static + layer0 output
        self.layer1 = nn.Conv2d(in_ch_1, channels[1], kernel_size=kernels[1],
                                padding=kernels[1]//2, bias=False)

        # Output layer: (7 features + bias + C_last) → 4 (RGBA)
        in_ch_2 = 8 + channels[1]
        self.layer2 = nn.Conv2d(in_ch_2, 4, kernel_size=kernels[2],
                                padding=kernels[2]//2, bias=False)

    def forward(self, static_features, layer0_input=None):
        # Layer 0: Use full 8D static features (includes bias)
        x0 = self.layer0(static_features)
        x0 = F.relu(x0)

        # Layer 1: Concatenate static + layer0 output
        x1_input = torch.cat([static_features, x0], dim=1)
        x1 = self.layer1(x1_input)
        x1 = F.relu(x1)

        # Layer 2: Concatenate static + layer1 output
        x2_input = torch.cat([static_features, x1], dim=1)
        output = self.layer2(x2_input)

        return torch.sigmoid(output)  # RGBA output [0,1]
```

**Training Configuration:**

```python
# Hyperparameters
kernels = [1, 3, 5]          # Per-layer kernel sizes
channels = [16, 8, 4]        # Per-layer output channels
learning_rate = 1e-3
batch_size = 16
epochs = 5000

# Training loop (standard PyTorch f32)
for epoch in range(epochs):
    for rgb_batch, depth_batch, target_batch in dataloader:
        # Compute static features
        static_feat = compute_static_features(rgb_batch, depth_batch)

        # Forward pass
        output = model(static_feat)
        loss = criterion(output, target_batch)

        # Backward pass
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
```

**Checkpoint Format:**

```python
torch.save({
    'state_dict': model.state_dict(),  # f32 weights
    'config': {
        'kernels': [1, 3, 5],
        'channels': [16, 8, 4],
        'features': ['R', 'G', 'B', 'D', 'uv.x', 'uv.y', 'sin10_x', 'bias']
    },
    'epoch': epoch,
    'loss': loss.item()
}, f'checkpoints/checkpoint_epoch_{epoch}.pth')
```

---

## Export Workflow

### Script: `training/export_cnn_v2_shader.py`

**Process:**
1. Load checkpoint (f32 PyTorch weights)
2. Extract layer configs (kernels, channels)
3. Quantize weights to float16: `weights_f16 = weights_f32.astype(np.float16)`
4. Generate WGSL shader per layer
5. Write to `workspaces/<workspace>/shaders/cnn_v2_*.wgsl`

**Example Generated Shader:**

```wgsl
// cnn_v2_layer_0.wgsl - Auto-generated from checkpoint_epoch_5000.pth

const KERNEL_SIZE: u32 = 1u;
const IN_CHANNELS: u32 = 8u;   // 7 features + bias
const OUT_CHANNELS: u32 = 16u;

// Weights quantized to float16 (stored as f32 in shader)
const weights: array<f32, 128> = array(
  0.123047, -0.089844, 0.234375, 0.456055, ...
);

@group(0) @binding(0) var static_features: texture_2d<u32>;
@group(0) @binding(1) var output_texture: texture_storage_2d<rgba32uint, write>;

@compute @workgroup_size(8, 8)
fn main(@builtin(global_invocation_id) id: vec3<u32>) {
  // Load static features (8D)
  let static_feat = get_static_features(vec2<i32>(id.xy));

  // Convolution (1×1 kernel = pointwise)
  var output: array<f32, OUT_CHANNELS>;
  for (var c: u32 = 0u; c < OUT_CHANNELS; c++) {
    var sum: f32 = 0.0;
    for (var k: u32 = 0u; k < IN_CHANNELS; k++) {
      sum += weights[c * IN_CHANNELS + k] * static_feat[k];
    }
    output[c] = max(0.0, sum);  // ReLU activation
  }

  // Pack and store (8×f16 per texel)
  textureStore(output_texture, vec2<i32>(id.xy), pack_f16x8(output));
}
```

**Float16 Quantization:**
- Training uses f32 throughout (PyTorch standard)
- Export converts to np.float16, then back to f32 for WGSL literals
- **Expected discrepancy:** <0.1% MSE (acceptable)
- Validation via `validate_cnn_v2.sh` compares outputs

---

## Validation Workflow

### Script: `scripts/validate_cnn_v2.sh`

**End-to-end pipeline:**
```bash
./scripts/validate_cnn_v2.sh checkpoints/checkpoint_epoch_5000.pth
```

**Steps automated:**
1. Export checkpoint → .wgsl shaders
2. Rebuild `cnn_test` tool
3. Process test images with CNN v2
4. Display input/output results

**Usage:**
```bash
# Basic usage
./scripts/validate_cnn_v2.sh checkpoint.pth

# Custom paths
./scripts/validate_cnn_v2.sh checkpoint.pth \
  -i my_test_images/ \
  -o results/ \
  -b build_release

# Skip rebuild (iterate on checkpoint only)
./scripts/validate_cnn_v2.sh checkpoint.pth --skip-build

# Skip export (iterate on test images only)
./scripts/validate_cnn_v2.sh checkpoint.pth --skip-export

# Show help
./scripts/validate_cnn_v2.sh --help
```

**Options:**
- `-b, --build-dir DIR` - Build directory (default: build)
- `-w, --workspace NAME` - Workspace name (default: main)
- `-i, --images DIR` - Test images directory (default: training/validation)
- `-o, --output DIR` - Output directory (default: validation_results)
- `--skip-build` - Use existing cnn_test binary
- `--skip-export` - Use existing .wgsl shaders
- `-h, --help` - Show full usage

**Output:**
- Input images: `<test_images_dir>/*.png`
- Output images: `<output_dir>/*_output.png`
- Opens results directory in system file browser

---

## Implementation Checklist

### Phase 1: Shaders (Core Infrastructure)

- [ ] `workspaces/main/shaders/cnn_v2_static.wgsl` - Static features compute
  - [ ] RGBD sampling from framebuffer
  - [ ] UV coordinate calculation
  - [ ] sin(10\*uv.x) computation
  - [ ] Bias dimension (constant 1.0)
  - [ ] Float16 packing via `pack2x16float()`
  - [ ] Output to `texture_storage_2d<rgba32uint>`

- [ ] `workspaces/main/shaders/cnn_v2_layer_template.wgsl` - Layer template
  - [ ] Static features unpacking
  - [ ] Previous layer unpacking (8×f16)
  - [ ] Convolution implementation (1×1, 3×3, 5×5)
  - [ ] ReLU activation
  - [ ] Output packing (8×f16)
  - [ ] Proper padding handling

### Phase 2: C++ Effect Class

- [ ] `src/gpu/effects/cnn_v2_effect.h` - Header
  - [ ] Class declaration inheriting from `PostProcessEffect`
  - [ ] Static features texture member
  - [ ] Layer textures vector
  - [ ] Pipeline and bind group members

- [ ] `src/gpu/effects/cnn_v2_effect.cc` - Implementation
  - [ ] Constructor: Load shaders, create textures
  - [ ] `init()`: Create pipelines, bind groups
  - [ ] `render()`: Multi-pass execution
    - [ ] Pass 0: Compute static features
    - [ ] Pass 1-N: CNN layers
    - [ ] Final: Composite to output
  - [ ] Proper resource cleanup

- [ ] Integration
  - [ ] Add to `src/gpu/demo_effects.h` includes
  - [ ] Add `cnn_v2_effect.cc` to `CMakeLists.txt` (headless + normal)
  - [ ] Add shaders to `workspaces/main/assets.txt`
  - [ ] Add to `src/tests/gpu/test_demo_effects.cc`

### Phase 3: Training Pipeline

- [ ] `training/train_cnn_v2.py` - Training script
  - [ ] Static feature extraction function
  - [ ] CNNv2 PyTorch model class
  - [ ] Patch-based dataloader
  - [ ] Training loop with checkpointing
  - [ ] Command-line argument parsing
  - [ ] Inference mode (ground truth generation)

- [ ] `training/export_cnn_v2_shader.py` - Export script
  - [ ] Checkpoint loading
  - [ ] Weight extraction and f16 quantization
  - [ ] Per-layer WGSL generation
  - [ ] File output to workspace shaders/
  - [ ] Metadata preservation

### Phase 4: Tools & Validation

- [ ] `scripts/validate_cnn_v2.sh` - End-to-end validation
  - [ ] Command-line argument parsing
  - [ ] Shader export orchestration
  - [ ] Build orchestration
  - [ ] Batch image processing
  - [ ] Results display

- [ ] `src/tools/cnn_test_main.cc` - Tool updates
  - [ ] Add `--cnn-version v2` flag
  - [ ] CNNv2Effect instantiation path
  - [ ] Static features pass execution
  - [ ] Multi-layer processing

### Phase 5: Documentation

- [ ] `doc/HOWTO.md` - Usage guide
  - [ ] Training section (CNN v2)
  - [ ] Export section
  - [ ] Validation section
  - [ ] Examples

- [ ] `README.md` - Project overview update
  - [ ] Mention CNN v2 capability

---

## File Structure

### New Files

```
# Shaders (generated by export script)
workspaces/main/shaders/cnn_v2_static.wgsl       # Static features compute
workspaces/main/shaders/cnn_v2_layer_0.wgsl      # Input layer (generated)
workspaces/main/shaders/cnn_v2_layer_1.wgsl      # Inner layer (generated)
workspaces/main/shaders/cnn_v2_layer_2.wgsl      # Output layer (generated)

# C++ implementation
src/gpu/effects/cnn_v2_effect.h                  # Effect class header
src/gpu/effects/cnn_v2_effect.cc                 # Effect implementation

# Python training/export
training/train_cnn_v2.py                         # Training script
training/export_cnn_v2_shader.py                 # Shader generator
training/validation/                             # Test images directory

# Scripts
scripts/validate_cnn_v2.sh                       # End-to-end validation

# Documentation
doc/CNN_V2.md                                    # This file
```

### Modified Files

```
src/gpu/demo_effects.h                           # Add CNNv2Effect include
CMakeLists.txt                                   # Add cnn_v2_effect.cc
workspaces/main/assets.txt                       # Add cnn_v2 shaders
workspaces/main/timeline.seq                     # Optional: add CNNv2Effect
src/tests/gpu/test_demo_effects.cc               # Add CNNv2 test case
src/tools/cnn_test_main.cc                       # Add --cnn-version v2
doc/HOWTO.md                                     # Add CNN v2 sections
TODO.md                                          # Add CNN v2 task
```

### Unchanged (v1 Preserved)

```
training/train_cnn.py                            # Original training
src/gpu/effects/cnn_effect.*                     # Original effect
workspaces/main/shaders/cnn_*.wgsl               # Original shaders
```

---

## Performance Characteristics

### Static Features Compute
- **Cost:** ~0.1ms @ 1080p
- **Frequency:** Once per frame
- **Operations:** sin(), texture sampling, packing

### CNN Layers (Example 3-layer)
- **Layer0 (1×1, 8→16):** ~0.3ms
- **Layer1 (3×3, 23→8):** ~0.8ms
- **Layer2 (5×5, 15→4):** ~1.2ms
- **Total:** ~2.4ms @ 1080p

### Memory Usage
- Static features: 1920×1080×8×2 = 33 MB (f16)
- Layer buffers: 1920×1080×16×2 = 66 MB (max 16 channels)
- Weights: ~6.4 KB (f16, in shader code)
- **Total GPU memory:** ~100 MB

---

## Size Budget

### CNN v1 vs v2

| Metric | v1 | v2 | Delta |
|--------|----|----|-------|
| Weights (count) | 800 | 3268 | +2468 |
| Storage (f32) | 3.2 KB | 13.1 KB | +9.9 KB |
| Storage (f16) | N/A | 6.5 KB | +6.5 KB |
| Shader code | ~500 lines | ~800 lines | +300 lines |

### Mitigation Strategies

**Reduce channels:**
- [16,8,4] → [8,4,4] saves ~50% weights
- [16,8,4] → [4,4,4] saves ~60% weights

**Smaller kernels:**
- [1,3,5] → [1,3,3] saves ~30% weights
- [1,3,5] → [1,1,3] saves ~50% weights

**Quantization:**
- int8 weights: saves 75% (requires QAT training)
- 4-bit weights: saves 87.5% (extreme, needs research)

**Target:** Keep CNN v2 under 10 KB for 64k demo constraint

---

## Future Extensions

### More Features (uint8 Packing)

```wgsl
// 16 uint8 features per texel (texture_storage_2d<rgba8unorm>)
// [R, G, B, D, uv.x, uv.y, sin10.x, sin10.y,
//  sin20.x, sin20.y, dx, dy, gray_mip1, gray_mip2, variance, bias]
```
- Trade precision for quantity
- Requires quantization-aware training

### Temporal Features

- Previous frame RGBA (motion awareness)
- Optical flow vectors
- Requires multi-frame buffer

### Learned Position Encodings

- Replace hand-crafted sin(10\*uv) with learned embeddings
- Requires separate embedding network
- Similar to NeRF position encoding

### Dynamic Architecture

- Runtime kernel size selection based on scene
- Conditional layer execution (skip connections)
- Layer pruning for performance

---

## References

- **v1 Implementation:** `src/gpu/effects/cnn_effect.*`
- **Training Guide:** `doc/HOWTO.md` (CNN Training section)
- **Test Tool:** `doc/CNN_TEST_TOOL.md`
- **Shader System:** `doc/SEQUENCE.md`
- **Size Measurement:** `doc/SIZE_MEASUREMENT.md`

---

## Appendix: Design Decisions

### Why Bias as Static Feature?

**Alternatives considered:**
1. Separate bias array per layer (Option B)
2. Bias as static feature = 1.0 (Option A, chosen)

**Decision rationale:**
- Simpler shader code (fewer bindings)
- Standard NN formulation (augmented input)
- Saves 56-112 bytes per model
- 7 features sufficient for v1 implementation
- Can extend to uint8 packing if >7 features needed

### Why Float16 for Weights?

**Alternatives considered:**
1. Keep f32 (larger, more accurate)
2. Use f16 (smaller, GPU-native)
3. Use int8 (smallest, needs QAT)

**Decision rationale:**
- f16 saves 50% vs f32 (critical for 64k target)
- GPU-native support (pack2x16float in WGSL)
- <0.1% accuracy loss (acceptable)
- Simpler than int8 quantization

### Why Multi-Frequency Position Encoding?

**Inspiration:** NeRF (Neural Radiance Fields)

**Benefits:**
- Helps network learn high-frequency details
- Better than raw UV coordinates
- Small footprint (1D per frequency)

**Future:** Add sin(20\*uv), sin(40\*uv) if >7 features available

---

**Document Version:** 1.0
**Last Updated:** 2026-02-12
**Status:** Design approved, ready for implementation