From eaf0bd855306e70ca03f2d6579b4d6551aff6482 Mon Sep 17 00:00:00 2001 From: skal Date: Thu, 12 Feb 2026 12:11:53 +0100 Subject: TODO: 8-bit weight quantization for 2× size reduction MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - 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) --- training/train_cnn_v2.py | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) (limited to 'training/train_cnn_v2.py') 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__() -- cgit v1.2.3