status

Some strage trouble, for a while, I can't upload output feature data,
please run infere_ovinoXX.X.py under venv to generate *.npy data

code

infere_ovino21.4.py
a sample code to perform inference and extract feature, based on "open-vino21.1"

infere_ovino22.1.py
a sample code to perform inference and extract feature, based on "open-vino22.1"

read_npy.py
a sample code to compare output features which are saved in .npy

model

model is int8-Alex model
int8 was done in onnx-runtime and nncf

onnx-runtime: model.test-int8-Alex.onnx-rt-int8.cut.onnx
nncf-package: model.test-int8-Alex.ovino-nncrf-int8.onnx

IR conversion was done in each openvino version:
ovino21.4_mo.cut
ovino22.1_mo.cut

generated output feature data

save_np_*,npy

saved output features via openvino21, and 22

save_np_ vino-version _ sample-image _ sample-model .npy

e.g.
save_np_ovino21.4_Accordion-842.jpg_model.test-int8-Alex.ovino-nncrf-int8.ovino21.4_mo.cut.xml.npy
save_np_ovino22.1_Accordion-842.jpg_model.test-int8-Alex.ovino-nncrf-int8.ovino22.1_mo.cut.xml.npy

the difference we observe b/w openvino version and onnx-op layers is as follows

for instance, usig "Accordion-842.jpg"
difference b/w vino21-22 using fake-quantize (nncf-quantization) :
 min 0.0
 max 0.0
difference b/w vino21-22 using quant/dequant-linear (onnx-rt) :
 min -0.03369951
 max +0.03252697

A model w/ fake-quantize layer outputs the same values b/w 21 and 22
A model w/ quant/dequant-linear layer outputs different values b/w 21 and 22