Help with finn compilation

[surajkarki66]

Thank you for the help. I would like to ask one question before closing this issue. Currently, I am working on an FPGA implementation of Yolo v8. After performing the QAT, can we able to compile the model to get a bit stream file using the standard FINN documentation, or do we have to modify the standard FINN library to work with this? I have read your research paper on multi-object tracking, and I am trying to implement it myself. Please help.

[mdanilow]

Hello, the changes I've made to FINN to enable yolov8 deployment are still currently being merged into standard FINN library. Check the "yolov8_dev" branch of my FINN fork: https://github.com/mdanilow/finn/tree/yolov8_dev. In notebooks/experiments/yolov8 directory there is a build_yolov8.py script with custom streamlining, floorplanning and hw conversion steps.

• To generate an onnx file from your trained model and, use brevitas.export.export_qonnx function. You can inject it in the test.py after the get_model() function so you can evaluate the model that you are exporting and be sure that the model performs the way you expect.

• To add a preprocessing step (that divides the uint8 input by 255) into the hardware pipeline you can follow the official cnv_end2end notebook: https://github.com/Xilinx/finn/blob/main/notebooks/end2end_example/bnn-pynq/cnv_end2end_example.ipynb.

• If you would like to take look at a dedicated PYNQ driver or play around with my deployment of yolov8 on alveo u55c, here is my PR for finn_exaples repo: Pre-built yolov8 demo for Alveo u55c Xilinx/finn-examples#113

Good luck

[surajkarki66]

Thank You for the help.

[mdanilow]

Hello, I've added an export.py script that spits out an .onnx with QONNX "Quant" nodes. Also, on my yolov8_dev fork, in the build_yolov8.py script I've added a tidy_up step that accepts this .onnx and adds a proper preprocessing.

Also, my changes were merged into the dev branch of the official FINN repo, so it should be possible now to build with my "build_yolov8.py" script from there.

[surajkarki66]

Thank you so much for the update.

[surajkarki66]

I am pretty excited to implement it.

[surajkarki66]

One thing!, Could you tell me which version of Vivado have you used for the FINN compilation?

[mdanilow]

One thing!, Could you tell me which version of Vivado have you used for the FINN compilation?

2023.2

[surajkarki66]

i have installed latest Vitis HLS and Vivado which is 2025.1
Will it work with these one also?

[surajkarki66]

Hello, help me. I am trying to compile my exported yolov8 model, and I am getting the error given in the figure

[surajkarki66]

I am following every step you have told me.

[mdanilow]

Most likely not all nodes in the graph are suitable for hw conversion, inspect the <your_model>_step_convert_to_hw_layers.onnx graph from your build directory. You can do it by netron file.onnx