PYNQ + Vitis AI Libraries


I am trying to deploy the SSD MobileNet v2 trained with the COCO dataset at PYNQ. I am using Jupyter notebooks.
It is possible to use Vitis AI libraries (Vitis-AI/tools/Vitis-AI-Library at master · Xilinx/Vitis-AI · GitHub) to reimplement the post-processing? Have these libraries been ported to Python?

Thank you for the support.

These libs run on the DPU.
You can use the DPU with PYNQ, but currently only up to 1.3.2.


Hello @cathalmccabe,

Thank you for your response.

Indeed I am already using that library. However, I think that it only provides the overlay to communicate with the DPU. However, the SSD MobileNet v2 is not fully compatible with the DPU, I had to remove the pre-processing and post-processing layers, and then, now, I have to reimplement them.
The pre-processing is easy to implement, and the examples already provide a framework for it. Although, post-processing requires decoding the confidence scores and the box encodings. The libraries in the Vitis AI Libraries, namely this one Vitis-AI/tools/Vitis-AI-Library/xnnpp/src/tfssd at master · Xilinx/Vitis-AI · GitHub, already provide this implementation, and I am avoiding translating it o Python.

I am trying to bind the libraries, but I think that the Vitis Ai 1.4 already provides this binding.