Xie Lin 12699e7952 Tflm: Add per-channel quantization support to FULLY_CONNECTED operator ... This commit fixes the inference accuracy issue in tflite-micro by adding per-channel quantization support to the FULLY_CONNECTED operator. Root Cause: - The fork only supported per-tensor quantization (single scale/zero_point for entire weight tensor) - Upstream added per-channel quantization (independent scale/zero_point per output channel) which provides better accuracy Changes: 1. tensorflow/lite/micro/kernels/fully_connected.h - Added is_per_channel flag - Added per_channel_output_multiplier array - Added per_channel_output_shift array 2. tensorflow/lite/micro/kernels/fully_connected_common.cc - Updated quantization parameter calculation to support per-channel 3. tensorflow/lite/micro/kernels/fully_connected.cc - Added conditional execution path for per-channel quantization - Preserved TFLITE_MODEL_COMPILER customization Signed-off-by: Xie Lin <xielin3@xiaomi.com>		2026-04-20 14:48:10 +08:00
..
lite	Tflm: Add per-channel quantization support to FULLY_CONNECTED operator	2026-04-20 14:48:10 +08:00
BUILD	More cleanup of the BUILD files. (#9)	2021-04-17 06:44:37 +00:00
extra_rules.bzl	Download Cadence Nature DSP from GitHub (#2241)	2023-10-05 17:06:55 +00:00
workspace.bzl	Update Flatbuffer version to 2.0.6 and generate schema local. (#1374)	2022-08-25 18:01:35 +00:00