wrong outputs order

[YeWenxuan64]

my model's outputs are out0[1,4200,4], out1[1,4200,2], out2[1,4200,10] which mean their sizes are 16800, 8400, 42000 .
but after inference, the outputs are 16800, 42000, 8400.

quantized_model.zip

[quic-zhanweiw]

How did you get the info that the model's outputs are out0[1,4200,4], out1[1,4200,2], out2[1,4200,10]?

[YeWenxuan64]

How did you get the info that the model's outputs are out0[1,4200,4], out1[1,4200,2], out2[1,4200,10]?

use Netron app or use onnxruntime inference the origin onnx model

[quic-zhanweiw]

You need check the outputs for the QNN context binary model, not the ONNX model.

[YeWenxuan64]

how the QNN context binary model change the output order when convert onnx model to it?

[quic-zhanweiw]

Can you share the original pytorch or onnx model to us? We can check it.

[YeWenxuan64]

my model's outputs are out0[1,4200,4], out1[1,4200,2], out2[1,4200,10] which mean their sizes are 16800, 8400, 42000 . but after inference, the outputs are 16800, 42000, 8400.

[quantized_model.zip](https://github.com/user-attachments/files/23481545/quantized_model.zip)

here，models in zip

[quic-zhanweiw]

It seems the while ONNX model is converted to QNN, the order has been changed.

[YeWenxuan64]

interesting…… how could it happened? This only happened in one of my models.

[tim202503]

hi 叶文轩 (@YeWenxuan64) ,

We will check it further. Can you also create a salesforce case in https://qualcomm-cdmatech-support.lightning.force.com/ to track this issue?

thanks,
brs

[YeWenxuan64]

hi @YeWenxuan64 ,

We will check it further. Can you also create a salesforce case in https://qualcomm-cdmatech-support.lightning.force.com/ to track this issue?

thanks, brs

i can not login it, maybe needs staff account or permission😭

[quic-zhanweiw]

We‘re working on this issue.

[quic-zhanweiw]

We support get input and output names now.
https://github.com/quic/ai-engine-direct-helper/blob/main/samples/python/real_esrgan_x4plus/real_esrgan_x4plus.py#L137

For the outputs order issue, we're still working on it.

[YeWenxuan64]

Mystery solved, the QAIRT-SDK's qairt-converter inexplicably changed my model's output order (and sometimes it will change model's input order too...) so annoying!!!!😡

[quic-zhanweiw]

叶文轩 (@YeWenxuan64) Thank you for your feedback and patience; we will strive to resolve this issue as soon as possible.