CNN v2: Use alpha channel for p3 depth feature + layer visualization

Training changes (train_cnn_v2.py):
- p3 now uses target image alpha channel (depth proxy for 2D images)
- Default changed from 0.0 → 1.0 (far plane semantics)
- Both PatchDataset and ImagePairDataset updated

Test tools (cnn_test.cc):
- New load_depth_from_alpha() extracts PNG alpha → p3 texture
- Fixed bind group layout: use UnfilterableFloat for R32Float depth
- Added --save-intermediates support for CNN v2:
  * Each layer_N.png shows 4 channels horizontally (1812×345 grayscale)
  * layers_composite.png stacks all layers vertically (1812×1380)
  * static_features.png shows 4 feature channels horizontally
- Per-channel visualization enables debugging layer-by-layer differences

HTML tool (index.html):
- Extract alpha channel from input image → depth texture
- Matches training data distribution for validation

Note: Current weights trained with p3=0 are now mismatched. Both tools
use p3=alpha consistently, so outputs remain comparable for debugging.
Retrain required for optimal quality.

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

1 files changed, 238 insertions, 2 deletions

diff --git a/tools/cnn_test.cc b/tools/cnn_test.cc
index 5823110..3fad2ff 100644
--- a/tools/cnn_test.cc
+++ b/tools/cnn_test.cc
@@ -169,6 +169,66 @@ static WGPUTexture load_texture(WGPUDevice device, WGPUQueue queue,
   return texture;
 }
 
+// Load PNG alpha channel as depth texture (or 1.0 if no alpha)
+static WGPUTexture load_depth_from_alpha(WGPUDevice device, WGPUQueue queue,
+                                          const char* path, int width,
+                                          int height) {
+  int w, h, channels;
+  uint8_t* data = stbi_load(path, &w, &h, &channels, 4);
+  if (!data || w != width || h != height) {
+    fprintf(stderr, "Error: failed to load depth from '%s'\n", path);
+    if (data) stbi_image_free(data);
+    return nullptr;
+  }
+
+  // Extract alpha channel (or use 1.0 if original was RGB)
+  std::vector<float> depth_data(width * height);
+  bool has_alpha = (channels == 4);
+  for (int i = 0; i < width * height; ++i) {
+    // Alpha is in data[i*4+3] (0-255), convert to float [0, 1]
+    // If no alpha channel, default to 1.0 (far plane)
+    depth_data[i] = has_alpha ? (data[i * 4 + 3] / 255.0f) : 1.0f;
+  }
+  stbi_image_free(data);
+
+  // Create R32Float depth texture
+  const WGPUTextureDescriptor depth_desc = {
+      .usage = WGPUTextureUsage_TextureBinding | WGPUTextureUsage_CopyDst,
+      .dimension = WGPUTextureDimension_2D,
+      .size = {static_cast<uint32_t>(width), static_cast<uint32_t>(height), 1},
+      .format = WGPUTextureFormat_R32Float,
+      .mipLevelCount = 1,
+      .sampleCount = 1,
+  };
+  WGPUTexture depth_texture = wgpuDeviceCreateTexture(device, &depth_desc);
+  if (!depth_texture) {
+    fprintf(stderr, "Error: failed to create depth texture\n");
+    return nullptr;
+  }
+
+  // Write depth data
+  const WGPUTexelCopyTextureInfo dst = {
+      .texture = depth_texture,
+      .mipLevel = 0
+  };
+  const WGPUTexelCopyBufferLayout layout = {
+      .bytesPerRow = static_cast<uint32_t>(width * sizeof(float)),
+      .rowsPerImage = static_cast<uint32_t>(height)
+  };
+  const WGPUExtent3D size = {
+      static_cast<uint32_t>(width),
+      static_cast<uint32_t>(height),
+      1
+  };
+  wgpuQueueWriteTexture(queue, &dst, depth_data.data(),
+                        depth_data.size() * sizeof(float), &layout, &size);
+
+  printf("Loaded depth from alpha: %dx%d (%s alpha)\n", width, height,
+         has_alpha ? "has" : "no");
+
+  return depth_texture;
+}
+
 // Create CNN render pipeline (5 bindings)
 // Takes both intermediate format (RGBA16Float) and final format (BGRA8Unorm)
 static WGPURenderPipeline create_cnn_pipeline(WGPUDevice device,
@@ -245,6 +305,57 @@ static bool save_png(const char* path, const std::vector<uint8_t>& pixels,
   return true;
 }
 
+// Create horizontal grayscale composite of layer outputs
+// Each layer is already 4x wide (showing 4 channels), stack them vertically
+static bool save_layer_composite(const char* dir, int width, int height, int num_layers) {
+  // Each layer PNG is already 4x wide with 4 channels side-by-side
+  int layer_width = width * 4;
+
+  // Load all layer images (they're already grayscale)
+  std::vector<std::vector<uint8_t>> layers(num_layers);
+  for (int i = 0; i < num_layers; ++i) {
+    char path[512];
+    snprintf(path, sizeof(path), "%s/layer_%d.png", dir, i);
+
+    int w, h, channels;
+    uint8_t* data = stbi_load(path, &w, &h, &channels, 1); // Load as grayscale
+    if (!data || w != layer_width || h != height) {
+      if (data) stbi_image_free(data);
+      fprintf(stderr, "Warning: failed to load layer %d for composite (expected %dx%d, got %dx%d)\n",
+              i, layer_width, height, w, h);
+      return false;
+    }
+
+    layers[i].assign(data, data + (layer_width * height));
+    stbi_image_free(data);
+  }
+
+  // Stack layers vertically
+  int composite_height = height * num_layers;
+  std::vector<uint8_t> composite(layer_width * composite_height);
+
+  for (int layer = 0; layer < num_layers; ++layer) {
+    for (int y = 0; y < height; ++y) {
+      int src_row_offset = y * layer_width;
+      int dst_row_offset = (layer * height + y) * layer_width;
+      memcpy(&composite[dst_row_offset], &layers[layer][src_row_offset], layer_width);
+    }
+  }
+
+  // Save as grayscale PNG (stacked vertically)
+  char composite_path[512];
+  snprintf(composite_path, sizeof(composite_path), "%s/layers_composite.png", dir);
+  if (!stbi_write_png(composite_path, layer_width, composite_height, 1,
+                      composite.data(), layer_width)) {
+    fprintf(stderr, "Error: failed to write composite PNG\n");
+    return false;
+  }
+
+  printf("Saved layer composite to '%s' (%dx%d, 4 layers stacked vertically)\n",
+         composite_path, layer_width, composite_height);
+  return true;
+}
+
 // Save PPM output (fallback)
 static bool save_ppm(const char* path, const std::vector<uint8_t>& pixels,
                      int width, int height) {
@@ -282,6 +393,7 @@ struct CNNv2LayerParams {
   uint32_t weight_offset;
   uint32_t is_output_layer;
   float blend_amount;
+  uint32_t is_layer_0;
 };
 
 struct CNNv2StaticFeatureParams {
@@ -433,6 +545,41 @@ static std::vector<uint8_t> readback_rgba32uint_to_bgra8(
   return result;
 }
 
+// Read RGBA32Uint and create 4x wide grayscale composite (each channel side-by-side)
+static std::vector<uint8_t> readback_rgba32uint_to_composite(
+    WGPUDevice device, WGPUQueue queue, WGPUTexture texture,
+    int width, int height) {
+
+  // First get BGRA8 data
+  std::vector<uint8_t> bgra = readback_rgba32uint_to_bgra8(device, queue, texture, width, height);
+  if (bgra.empty()) return {};
+
+  // Create 4x wide grayscale image (one channel per horizontal strip)
+  int composite_width = width * 4;
+  std::vector<uint8_t> composite(composite_width * height);
+
+  for (int y = 0; y < height; ++y) {
+    for (int x = 0; x < width; ++x) {
+      int src_idx = (y * width + x) * 4;
+      uint8_t b = bgra[src_idx + 0];
+      uint8_t g = bgra[src_idx + 1];
+      uint8_t r = bgra[src_idx + 2];
+      uint8_t a = bgra[src_idx + 3];
+
+      // Convert each channel to grayscale luminance
+      auto to_gray = [](uint8_t val) -> uint8_t { return val; };
+
+      // Place each channel in its horizontal strip
+      composite[y * composite_width + (0 * width + x)] = to_gray(r); // Channel 0
+      composite[y * composite_width + (1 * width + x)] = to_gray(g); // Channel 1
+      composite[y * composite_width + (2 * width + x)] = to_gray(b); // Channel 2
+      composite[y * composite_width + (3 * width + x)] = to_gray(a); // Channel 3
+    }
+  }
+
+  return composite;
+}
+
 // Process image with CNN v2
 static bool process_cnn_v2(WGPUDevice device, WGPUQueue queue,
                            WGPUInstance instance, WGPUTexture input_texture,
@@ -523,6 +670,18 @@ static bool process_cnn_v2(WGPUDevice device, WGPUQueue queue,
   WGPUTextureView static_features_view =
       wgpuTextureCreateView(static_features_tex, nullptr);
 
+  // Load depth from input alpha channel (or 1.0 if no alpha)
+  WGPUTexture depth_texture =
+      load_depth_from_alpha(device, queue, args.input_path, width, height);
+  if (!depth_texture) {
+    wgpuTextureViewRelease(static_features_view);
+    wgpuTextureRelease(static_features_tex);
+    wgpuBufferRelease(weights_buffer);
+    wgpuTextureViewRelease(input_view);
+    return false;
+  }
+  WGPUTextureView depth_view = wgpuTextureCreateView(depth_texture, nullptr);
+
   // Create layer textures (ping-pong)
   WGPUTexture layer_textures[2] = {
       wgpuDeviceCreateTexture(device, &static_desc),
@@ -543,6 +702,8 @@ static bool process_cnn_v2(WGPUDevice device, WGPUQueue queue,
     fprintf(stderr, "Error: CNN v2 shaders not available\n");
     wgpuTextureViewRelease(static_features_view);
     wgpuTextureRelease(static_features_tex);
+    wgpuTextureViewRelease(depth_view);
+    wgpuTextureRelease(depth_texture);
     wgpuTextureViewRelease(layer_views[0]);
     wgpuTextureViewRelease(layer_views[1]);
     wgpuTextureRelease(layer_textures[0]);
@@ -600,7 +761,7 @@ static bool process_cnn_v2(WGPUDevice device, WGPUQueue queue,
 
   static_bgl_entries[3].binding = 3;
   static_bgl_entries[3].visibility = WGPUShaderStage_Compute;
-  static_bgl_entries[3].texture.sampleType = WGPUTextureSampleType_Float;
+  static_bgl_entries[3].texture.sampleType = WGPUTextureSampleType_UnfilterableFloat;
   static_bgl_entries[3].texture.viewDimension = WGPUTextureViewDimension_2D;
 
   static_bgl_entries[4].binding = 4;
@@ -651,7 +812,7 @@ static bool process_cnn_v2(WGPUDevice device, WGPUQueue queue,
   static_bg_entries[2].binding = 2;
   static_bg_entries[2].textureView = input_view;
   static_bg_entries[3].binding = 3;
-  static_bg_entries[3].textureView = input_view;  // Depth (use input)
+  static_bg_entries[3].textureView = depth_view;  // Depth from alpha channel (matches training)
   static_bg_entries[4].binding = 4;
   static_bg_entries[4].textureView = static_features_view;
   static_bg_entries[5].binding = 5;
@@ -769,6 +930,43 @@ static bool process_cnn_v2(WGPUDevice device, WGPUQueue queue,
   wgpuComputePassEncoderEnd(static_pass);
   wgpuComputePassEncoderRelease(static_pass);
 
+  // Save static features if requested
+  if (args.save_intermediates) {
+    // Submit and wait for static features to complete
+    WGPUCommandBuffer cmd = wgpuCommandEncoderFinish(encoder, nullptr);
+    wgpuQueueSubmit(queue, 1, &cmd);
+    wgpuCommandBufferRelease(cmd);
+    wgpuDevicePoll(device, true, nullptr);
+
+    // Create new encoder for layers
+    encoder = wgpuDeviceCreateCommandEncoder(device, nullptr);
+
+    char layer_path[512];
+    snprintf(layer_path, sizeof(layer_path), "%s/static_features.png",
+             args.save_intermediates);
+    printf("Saving static features to '%s'...\n", layer_path);
+
+    // Read back RGBA32Uint and create 8-channel grayscale composite
+    // Static features has 8 channels (packed as 4×u32), create 8x wide composite
+    std::vector<uint8_t> bgra = readback_rgba32uint_to_bgra8(
+        device, queue, static_features_tex, width, height);
+
+    if (!bgra.empty()) {
+      // Static features: 8 f16 values packed in 4×u32
+      // For now, just show first 4 channels (like layers)
+      // TODO: Show all 8 channels in 8x wide composite
+      std::vector<uint8_t> composite = readback_rgba32uint_to_composite(
+          device, queue, static_features_tex, width, height);
+      if (!composite.empty()) {
+        int composite_width = width * 4;
+        if (!stbi_write_png(layer_path, composite_width, height, 1,
+                            composite.data(), composite_width)) {
+          fprintf(stderr, "Error: failed to write static features PNG\n");
+        }
+      }
+    }
+  }
+
   // Pass 2-N: CNN layers
   for (size_t i = 0; i < layer_info.size(); ++i) {
     const CNNv2LayerInfo& info = layer_info[i];
@@ -785,6 +983,7 @@ static bool process_cnn_v2(WGPUDevice device, WGPUQueue queue,
     params.weight_offset = info.weight_offset;
     params.is_output_layer = (i == layer_info.size() - 1) ? 1 : 0;
     params.blend_amount = args.blend;
+    params.is_layer_0 = (i == 0) ? 1 : 0;
 
     wgpuQueueWriteBuffer(queue, layer_params_buffers[i], 0, &params,
                          sizeof(params));
@@ -831,6 +1030,36 @@ static bool process_cnn_v2(WGPUDevice device, WGPUQueue queue,
     wgpuComputePassEncoderEnd(layer_pass);
     wgpuComputePassEncoderRelease(layer_pass);
     wgpuBindGroupRelease(layer_bg);
+
+    // Save intermediate layer if requested
+    if (args.save_intermediates) {
+      // Submit and wait for layer to complete
+      WGPUCommandBuffer cmd = wgpuCommandEncoderFinish(encoder, nullptr);
+      wgpuQueueSubmit(queue, 1, &cmd);
+      wgpuCommandBufferRelease(cmd);
+      wgpuDevicePoll(device, true, nullptr);
+
+      // Create new encoder for next layer
+      encoder = wgpuDeviceCreateCommandEncoder(device, nullptr);
+
+      char layer_path[512];
+      snprintf(layer_path, sizeof(layer_path), "%s/layer_%zu.png",
+               args.save_intermediates, i);
+      printf("Saving intermediate layer %zu to '%s'...\n", i, layer_path);
+
+      // Read back RGBA32Uint and create 4-channel grayscale composite
+      WGPUTexture output_tex = layer_textures[(i + 1) % 2];
+      std::vector<uint8_t> composite = readback_rgba32uint_to_composite(
+          device, queue, output_tex, width, height);
+
+      if (!composite.empty()) {
+        int composite_width = width * 4;
+        if (!stbi_write_png(layer_path, composite_width, height, 1,
+                            composite.data(), composite_width)) {
+          fprintf(stderr, "Error: failed to write layer PNG\n");
+        }
+      }
+    }
   }
 
   WGPUCommandBuffer commands = wgpuCommandEncoderFinish(encoder, nullptr);
@@ -840,6 +1069,11 @@ static bool process_cnn_v2(WGPUDevice device, WGPUQueue queue,
 
   wgpuDevicePoll(device, true, nullptr);
 
+  // Create layer composite if intermediates were saved
+  if (args.save_intermediates) {
+    save_layer_composite(args.save_intermediates, width, height, layer_info.size());
+  }
+
   // Readback final result (from last layer's output texture)
   printf("Reading pixels from GPU...\n");
   size_t final_layer_idx = (layer_info.size()) % 2;
@@ -856,6 +1090,8 @@ static bool process_cnn_v2(WGPUDevice device, WGPUQueue queue,
     wgpuBufferRelease(static_params_buffer);
     wgpuTextureViewRelease(static_features_view);
     wgpuTextureRelease(static_features_tex);
+    wgpuTextureViewRelease(depth_view);
+    wgpuTextureRelease(depth_texture);
     wgpuTextureViewRelease(layer_views[0]);
     wgpuTextureViewRelease(layer_views[1]);
     wgpuTextureRelease(layer_textures[0]);