Drop in accuracy of Output between quantized Brevitas model and the streamlined Finn onnx model output.

[Arbiter-glitch]

Iam a Research student working on image processing. While using FINN, I was expecting the accuracy to be maintained between the trained brevitas model and the finn onnx model after streamlining. But even though I am getting the output there is a considerable drop in the image quality metrics. Is it an expected drop in accuracy or is it supposed to be the same. My files have been posted in earlier closed issues. here and here

[Update:] I have found that the ouput tensor varies from the first step: qonnx to finn conversion onwards. I dont get the same image quality from onnx execution like how it is in quantized brevitas software model. Is this normal? Iam having a 5dB drop in PSNR between software and Finn output.