Model and DPU fingerprint check fail - PyTorch flow for Vitis-AI

[Werterror]

Hi all,
I am following Vitis-AI tutorial "PyTorch flow for Vitis-AI":
https://github.com/Xilinx/Vitis-AI-Tutorials/blob/1.4/Design_Tutorials/09-mnist_pyt/README.md

I got through everything without a problem to the step where I try to launch model on zcu102, with commend:
python3 app_mt.py -m CNN_zcu102.xmodel
Then after pre-processing of images, when DPU is in use, error stating that there is a mismatch in model_fingerprint and dpu_fingerprint: 0x101000016010407 for model and 0x1000020f6014407 for dpu.

Tutorial that I am working on was written for version 1.4 of Vitis-AI while I belive I have image for 2.0.
If my understanding is correct and the problem is mismatching versions of DPU, how do I change the models DPU version to match one that I have on my ZCU102?

Screenshot of an error:

Edit:
In another thread that I found, states that editing fingerprint in json file could solve the problem,
but in both json files that I have don't contain variable named fingerprint,

1. meta.json in compiled_model:
   { "lib": "libvart-dpu-runner.so", "filename": "CNN_zcu102.xmodel", "kernel": [ "subgraph_CNN__CNN_Sequential_network__Conv2d_0__input_3" ], "target": "DPUCZDX8G_ISA1_B4096" }
2. quat_info.json in quant_model:
   { "param": { "CNN::network.0.weight":[8,7], "CNN::network.0.bias":[8,8], "CNN::network.3.weight":[8,8], "CNN::network.3.bias":[8,6], "CNN::network.6.weight":[8,8], "CNN::network.6.bias":[8,6], "CNN::network.9.weight":[8,7], "CNN::network.9.bias":[8,4] }, "output": { "CNN::input_0":[8,6], "CNN::CNN/Sequential[network]/ReLU[2]/input.7":[8,6], "CNN::CNN/Sequential[network]/ReLU[5]/input.13":[8,5], "CNN::CNN/Sequential[network]/ReLU[8]/input.19":[8,5], "CNN::CNN/Sequential[network]/Conv2d[9]/input":[8,4] }, "input": { }, "fast_finetuned":false, "bias_corrected":true }

Although in meta.json there is target DPU:
"target": "DPUCZDX8G_ISA1_B4096"
Would changing that to matching name on zcu102 solve the problem, if yes where can I find DPU name on zcu102?

[lishixlnx]

try
env XLNX_ENABLE_FINGERPRINT_CHECK=0 python3 app_mt.py -m CNN_zcu102.xmodel
to check if it works

[Werterror]

@lishixlnx thank you for reply,
but sadly it didnt worked either. Probably either of DPUs have to be changed so that they match each other.
Image with error:

[lishixlnx]

compile your model with the correct fingerprint.
find arch.json in your docker , /opt/vitis_ai/compiler/arch
change the content something like
{ "target": "DPUCZDX8G_ISA1_B4096"}
to
{"fingerprint":"0x0101000016010407"}
where 0x0101000016010407 should be your own dpu_fingerprint in above error message.

[Werterror]

Hi @lishixlnx,
Again thank you for your reply.
After changing target to fingerprint it still doesnt work,
first error occurs (DPU Fingerprint mismatch), also just to make sure I tried that twice.
The thing is that "meta.json" file that I am editing is created After compiling the model,
it might be more of a description file rather than setting file.

Whole project compressed: https://filetransfer.io/data-package/65crYaP8#link

[lishixlnx]

please notice, I mean arch.json , not meta.json.
please compile your model in the docker envrionment.

[Werterror]

Hi, @lishixlnx, I tried to look for "arch.json" but I can't find any file with such name in the project, that's why I included the files.
According to what you wrote I should edit that file before compilation of the model, that means right after quantization.
But after quantization only json file that I can find in the project is quant_info.json.
Also where else could I compile project if not in the docker?

[lishixlnx]

in document https://docs.xilinx.com/r/en-US/ug1414-vitis-ai/Compiling-for-DPU
https://docs.xilinx.com/r/en-US/ug1414-vitis-ai/VAI_C-Usage,
it only mentioned the way of compiling in docker and use the arch.json.
Please follow that standard way.

[qianglin-xlnx]

Hi @Werterror
Usually, when users built and integrated DPU IP, it will generate arch.json file, as show below. You can also check the DPU fingerprint on the target and create the arch.json file by yourself, just like @lishixlnx said. The arch.json file is used for model compiling. Different DPU IP will have different fingerprint.

https://docs.xilinx.com/r/2.0-English/ug1414-vitis-ai/Deep-Learning-Processor-Unit

BTW, could you run the model successfully now?

[Werterror]

Hi @qianglin-xlnx,
The thing is that arch.json didn't get created during Vits-AI flow hence there is nothing to be edited.
Also project that is subject of this discussion doesn't use arch.json file to compile its model.
I will see if I can figure out how to create one and use it with example project.
BR

[qianglin-xlnx]

Hi @qianglin-xlnx
You can create the arch.json. It's one of the inputs for the compiler, as shown in the compile.sh script below.
https://github.com/Xilinx/Vitis-AI-Tutorials/blob/1.4/Design_Tutorials/09-mnist_pyt/files/compile.sh#L20

[Werterror]

Hi @qianglin-xlnx,
I will try that solution next week and let you know about the results.
Thank you for all the resources.

[Werterror]

Sadly I dont have time to test this solution.
If you think its appropriate, close the issue and I will create new one if it doesnt solve the issue.
Best Regards