Nov 4 4 days ago 11:02 Drop QNN a16w8 segmentation variants; a8w8 matches their mIoU at half the latency Oct 20 19 days ago 12:22 Export the MobileNetV3 QNN variants as a8w8 (2.2x faster at the same mIoU) Oct 19 20 days ago 18:32 Export the ResNet QNN variants as a8w8 (2x faster at the same mIoU) Sep 25 20:17 QNN: re-export DeepLabV3-MobileNetV3 a16w8 with eps=1e-5 (66.55% -> 69.02% mIoU) 18:16 QNN: DeepLabV3-ResNet101 a16w8; list the ResNet QNN files in the README 14:30 QNN: DeepLabV3-ResNet50 a16w8, one context binary per Hexagon version Sep 24 12:58 QNN: DeepLabV3-MobileNetV3 a16w8, one context binary per Hexagon version 66.55% mIoU on VOC2012 val (200 images) against 66.56% for the XNNPACK int8
file, at 3.20 ms/inference on an S26 Ultra. Returns a class index per pixel,
(1,520,520) int32, instead of (1,21,520,520) logits. 
