diff options
Diffstat (limited to 'doc/HOWTO.md')
| -rw-r--r-- | doc/HOWTO.md | 52 |
1 files changed, 34 insertions, 18 deletions
diff --git a/doc/HOWTO.md b/doc/HOWTO.md index 7be5246..1ae1d94 100644 --- a/doc/HOWTO.md +++ b/doc/HOWTO.md @@ -132,36 +132,52 @@ Processes entire image with sliding window (matches WGSL): ### CNN v2 Training -Enhanced CNN with parametric static features (7D input: RGBD + UV + sin encoding + bias): +Enhanced CNN with parametric static features (7D input: RGBD + UV + sin encoding + bias). +**Complete Pipeline** (recommended): ```bash -# Train CNN v2 with default config (1×1, 3×3, 5×5 kernels, 16→8→4 channels) +# Train → Export → Build → Validate +./scripts/train_cnn_v2_full.sh +``` + +Config: 100 epochs, 3×3 kernels, 8→4→4 channels, patch-based (harris detector). +- Live progress with single-line update +- Validates all input images on final epoch +- Exports binary weights (storage buffer architecture) + +**Validation Only** (skip training): +```bash +# Use latest checkpoint +./scripts/train_cnn_v2_full.sh --validate + +# Use specific checkpoint +./scripts/train_cnn_v2_full.sh --validate checkpoints/checkpoint_epoch_50.pth +``` + +**Manual Training:** +```bash +# Default config ./training/train_cnn_v2.py \ - --input training/input/ --target training/output/ \ - --epochs 5000 --batch-size 16 \ - --checkpoint-every 1000 + --input training/input/ --target training/target_2/ \ + --epochs 100 --batch-size 16 --checkpoint-every 5 -# Custom architecture (smaller for size optimization) +# Custom architecture ./training/train_cnn_v2.py \ - --input training/input/ --target training/output/ \ - --kernel-sizes 1 3 3 --channels 8 4 4 \ + --input training/input/ --target training/target_2/ \ + --kernel-sizes 1 3 5 --channels 16 8 4 \ --epochs 5000 --batch-size 16 ``` -**Export shaders:** +**Export Binary Weights:** ```bash -./training/export_cnn_v2_shader.py checkpoints/checkpoint_epoch_5000.pth \ - --output-dir workspaces/main/shaders +./training/export_cnn_v2_weights.py checkpoints/checkpoint_epoch_100.pth \ + --output-weights workspaces/main/cnn_v2_weights.bin ``` -Generates `cnn_v2_layer_0.wgsl`, `cnn_v2_layer_1.wgsl`, `cnn_v2_layer_2.wgsl` with f16 weights. +Generates binary format: header + layer info + f16 weights (~3.2 KB for 3-layer model). +Storage buffer architecture allows dynamic layer count. -### CNN v2 Validation - -End-to-end testing: checkpoint → shaders → build → test images → results - -```bash -./scripts/validate_cnn_v2.sh checkpoints/checkpoint_epoch_5000.pth +**TODO:** 8-bit quantization for 2× size reduction (~1.6 KB). Requires quantization-aware training (QAT). # Options: # -i DIR Test images directory (default: training/validation) |