diff options
Diffstat (limited to 'TODO.md')
| -rw-r--r-- | TODO.md | 26 |
1 files changed, 9 insertions, 17 deletions
@@ -31,26 +31,18 @@ Enhanced CNN post-processing with multi-dimensional feature inputs. **Design:** `doc/CNN_V2.md` **Status:** -- ✅ Phase 1-5: All implementation phases complete -- ✅ Storage buffer architecture (dynamic layer count support) -- ✅ Binary weight format (header + layer info + f16 weights) -- ✅ Training pipeline with live progress display -- ✅ Complete validation tooling (`train_cnn_v2_full.sh --validate`) +- ✅ Full implementation complete and validated +- ✅ Binary weight loading fixed (FATAL_CHECK inversion bug) +- ✅ Training pipeline: 100 epochs, 3×3 kernels, patch-based - ✅ All tests passing (36/36) -**Features:** -- 7D static feature input (RGBD + UV + sin encoding + bias) -- Storage buffer weights (~3.2 KB for 3-layer, 8→4→4 config) -- Dynamic layer count (not hardcoded) -- Single compute shader with per-layer params -- Patch-based training (harris detector, 32×32 patches) -- Fast training config: 100 epochs, 3×3 kernels +**Specs:** +- 7D static features (RGBD + UV + sin + bias) +- Storage buffer weights (~3.2 KB, 8→4→4 channels) +- Dynamic layer count, per-layer params +- <10 KB target achieved -**Performance:** -- Storage buffer overhead: ~10-20% vs constants (negligible @ 60fps) -- Target achieved: <10 KB for 64k demo constraint - -**TODO:** 8-bit quantization for 2× size reduction (~1.6 KB). Requires QAT. +**TODO:** 8-bit quantization (2× reduction, needs QAT). --- |