Python inference gives always same result on quantized model regardless of input

[dkwiatko]

Hello, i described my situation here

https://github.com/openvinotoolkit/openvino/issues/4317 

In general my inference code in Python "just works", I get result corresponding to input.

However after default pot quantization, inference gives always the same result regardless input I pass to executable network.

Have anyone spotted similar behavior?  

[Iffa_Intel · Accepted solution]

Greetings,

I've checked the POT with mobilenet-v2-pytorch and tested the ori model, converted FP32 model and Quantized model with the benchmark_app. Each produces a different performance.

For Ori model:

Latency: 18.90 ms

Throughput: 191.67 FPS

For FP32 model:

Latency: 13.02 ms

Throughput: 299.82 FPS

For Quantized model:

Latency: 9.12 ms

Throughput: 456.67 FPS

Besides, I tested inferencing on the Quantized model and give different input and the result is good so far.

You may refer to my attachment for further detail. (Download them)

I had attached the config files together.

Hope this helps!

Sincerely,

Iffa

[Iffa_Intel]

Greetings,

Intel will no longer monitor this thread since we have provided a solution. If you need any additional information from Intel, please submit a new question. 

Sincerely,

Iffa