Description
使用的fp32模型已打包如下
量化使用的数据集WIDERFACE中的WIDER Face Testing Images中的16--Award_Ceremony:
下载地址:https://drive.google.com/file/d/1HIfDbVEWKmsYKJZm4lchTBDLW5N7dY5T/view
详细描述
依次使用onnx2ncnn → ncnnoptimize → ncnn2table →ncnn2int8
最终生成的int8模型大小与fp32模型的大小对比如下,int8模型与预期大小不符
-rw-rw-r-- 1 tjy tjy 3.9M Jan 24 13:38 hand_landmark_lite.bin
-rw-rw-r-- 1 tjy tjy 3.4M Jan 24 13:38 hand_landmark_lite-int8.bin
-rw-rw-r-- 1 tjy tjy 34K Jan 24 13:38 hand_landmark_lite-int8.param
-rw-rw-r-- 1 tjy tjy 3.9M Jan 24 13:38 hand_landmark_lite-opt.bin
-rw-rw-r-- 1 tjy tjy 34K Jan 24 13:38 hand_landmark_lite-opt.param
-rw-rw-r-- 1 tjy tjy 55K Jan 24 13:38 hand_landmark_lite.param
-rw-rw-r-- 1 tjy tjy 4.0M Jan 24 13:38 hand_landmark_lite-sim.onnx
-rw-rw-r-- 1 tjy tjy 133K Jan 24 13:38 hand_landmark_lite.table
详细命令:
onnx2ncnn hand_landmark_lite-sim.onnx
hand_landmark_lite.param
hand_landmark_lite.bin
ncnnoptimize hand_landmark_lite.param
hand_landmark_lite.bin
hand_landmark_lite-opt.param
hand_landmark_lite-opt.bin
0
ncnn2table hand_landmark_lite-opt.param
hand_landmark_lite-opt.bin
mydataset.txt
hand_landmark_lite.table
mean=[0,0,0]
norm=[0.003922,0.003922,0.003922]
norm=[0.003922,0.003922,0.003922]
shape=[224,224,3]
pixel=RGB
thread=8
method=aciq
ncnn2int8 hand_landmark_lite-opt.param
hand_landmark_lite-opt.bin
hand_landmark_lite-int8.param
hand_landmark_lite-int8.bin
hand_landmark_lite.table
Activity