search

Xilinx / QNN-MO-PYNQ

BSD 3-Clause "New" or "Revised" License
242 stars 114 forks source link

Tinier YOLO processing flow #9

Closed joyliu37 closed 6 years ago

joyliu37 commented 6 years ago

Hi,

I am wondering if this Tinier YOLO design have the same flow with Tincy YOLO shown in Xilinx website? https://forums.xilinx.com/t5/Xcell-Daily-Blog/Tincy-YOLO-a-real-time-low-latency-low-power-object-detection/ba-p/815840

giuliogamba commented 6 years ago

Hi,

yes, it's the same but the data reported in the Xcell blog are about the implementation on a different board, namely the Ultra96.

joyliu37 commented 6 years ago

Thanks for your quick reply. Did you implement all the optimization? I found in the code you executing the first and last layer in python instead of the NEON acceleration. And maxpool is still in the architecture. Moreover, could tell me the compressed model's mAP and on PASCAL VOC. Thank you so much!

giuliogamba commented 6 years ago

It's the same model used on the other board, with 48.5 mAP on Pascal VOC. The optimization using NEON are not included in the current release, we do use a Python backend to evaluate first and last layer.