TODO: 8-bit weight quantization for 2× size reduction

- Add QAT (quantization-aware training) notes
- Requires training with fake quantization
- Target: ~1.6 KB weights (vs 3.2 KB f16)
- Shader unpacking needs adaptation (4× u8 per u32)

2 files changed, 9 insertions, 1 deletions

diff --git a/training/export_cnn_v2_weights.py b/training/export_cnn_v2_weights.py
index e3d1724..723f572 100755
--- a/training/export_cnn_v2_weights.py
+++ b/training/export_cnn_v2_weights.py
@@ -94,6 +94,8 @@ def export_weights_binary(checkpoint_path, output_path):
     weight_offset += len(layer2_flat)
 
     # Convert to f16
+    # TODO: Use 8-bit quantization for 2× size reduction
+    # Requires quantization-aware training (QAT) to maintain accuracy
     all_weights_f16 = np.array(all_weights, dtype=np.float16)
 
     # Pack f16 pairs into u32 for storage buffer
diff --git a/training/train_cnn_v2.py b/training/train_cnn_v2.py
index 5c93f20..8cac51a 100755
--- a/training/train_cnn_v2.py
+++ b/training/train_cnn_v2.py
@@ -52,7 +52,13 @@ def compute_static_features(rgb, depth=None):
 
 
 class CNNv2(nn.Module):
-    """CNN v2 with parametric static features."""
+    """CNN v2 with parametric static features.
+
+    TODO: Add quantization-aware training (QAT) for 8-bit weights
+    - Use torch.quantization.QuantStub/DeQuantStub
+    - Train with fake quantization to adapt to 8-bit precision
+    - Target: ~1.6 KB weights (vs 3.2 KB with f16)
+    """
 
     def __init__(self, kernels=[1, 3, 5], channels=[16, 8, 4]):
         super().__init__()