01-03-2019 01:35 AM - edited 01-07-2019 10:11 PM
I want to run my own training YOLO V3 model on ZCU102 or Ultra96 boards, but there is no detailed introduction in the DNNK user guide. I have some questions at present: 1. There are some layers in Yolo V3 that Caffe does not support. Is the model trained with caffe? Can you provide detailed steps?
01-23-2019 09:14 AM
@lixun123YOLOv3 is trained using DarkNet and then deployed using a DarkNet to Caffe model conversion tool. To the best of my knowledge, the layer support for YOLOv3 is already in place and it can be directly deployed using this flow. I would suggest that this discussion might be best driven outside of the forums. I will send you a PM and we can try to connect you with the right team members.
02-25-2019 05:09 PM
Also interested in doing this:
1- Training from Darknet (yolov3 or yolov3-tiny)
2- Darknet to Caffe
3- Caffe to Xilinx
Please also send me information.
03-06-2019 05:45 AM
@qhall Is there any method to run yolov3 on ZCU102 board with high throughput and speed(like 30fps of yolov3-tiny)? I also wonder if you know how to convert the yolov3 from darknet to caffe while the transformed model can be quantized properly by the Decent tool. I am really confused about this conversion step. And how to solve the quantization of layers which caffe does not support?
03-06-2019 08:52 AM