The difference between openvino results and quantized openvino results in Real-ESRGAN

[jchin]

I converted the REAL-ESRGAN model into openvino (2021.4) and executed it. Although it is not perfect but it can run like below

the left picture is original picture with bicubic interpolation, left is I run real-esrgan with openvino. The IR file is in IRv10 which is converted from onnx

But after I use POT(Post-Training Optimization Toolkit) to convert the IR file from fp32 to the int8, the result is totally wrong; there are several minus float numbers in the output  

Could anyone tell me why the POT run differently with the expectation? Thanks a lot

Here is my inference processing code

The IR files are here

https://drive.google.com/file/d/1CL8cHWEC78W6-Vbu3bAfmE1wpxhPJMBY/view?usp=sharing

https://drive.google.com/file/d/1J3OQyEJdMlRHi2z4bBhNi4WfI1PD3UII/view?usp=sharing

[Iffa_Intel]

Greetings,

Compressing a full precision model into a smaller size format definitely has its own drawback.

The inferencing time would be faster but the trade-off is the accuracy.

If your use case requires to be accurate, such as clinical related results, it is not recommended to shrink the model as you need to bear with a less accurate prediction. Meanwhile, if your use case requires fast results without the need for prediction to be really accurate, then it is suitable to use a smaller size format.

This might help you to understand better.

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