UPGRADE YOUR BROWSER

We have detected your current browser version is not the latest one. Xilinx.com uses the latest web technologies to bring you the best online experience possible. Please upgrade to a Xilinx.com supported browser:Chrome, Firefox, Internet Explorer 11, Safari. Thank you!

cancel
Showing results for 
Search instead for 
Did you mean: 
Visitor avisakio
Visitor
291 Views
Registered: ‎03-15-2019

problem about yolov3 on ZCU104 with dnndk

Problem about deploying yolov3 on ZCU104 with dnndk
 
When I use the weights file trained on COCO offered in the repo, everything is good. I can get the right detection result on my ZCU104 board.
But when I use the weights file trained on my own dataset, following the same steps (convert to caffe model, run decent, run dnnc, test on board), I can't get right results on my board. I got 10000+ boxes size(as showed in the screenshot), which is not normal, and the process just stay there without showing anything else.
I tested the caffe model converted from darknet model, the result is ok. So the problem should be in steps of decent or dnnc.
The commands I used for decent and dnnc(the same as the tutorials):
decent quantize -model 2_model_for_quantize/v3.prototxt \ -weights 2_model_for_quantize/v3.caffemodel \ -gpu 0 \ -sigmoided_layers layer81-conv,layer93-conv,layer105-conv \ -output_dir 3_model_after_quantize \ -method 1
dnnc-dpu1.3.0 --prototxt=3_model_after_quantize/deploy.prototxt \ --caffemodel=3_model_after_quantize/deploy.caffemodel \ --dpu=4096FA \ --cpu_arch=arm64 --output_dir=4_model_elf \ --net_name=yolo --mode=normal --save_kernel
Can someone help me and tell me what happened, or give me some advice about what I should do?
ApplicationFrameHost_2019-10-09_21-38-21.png
0 Kudos
4 Replies
Xilinx Employee
Xilinx Employee
242 Views
Registered: ‎03-21-2008

Re: problem about yolov3 on ZCU104 with dnndk

Are you using the same anchor points for the training with your dataset?

0 Kudos
Visitor avisakio
Visitor
188 Views
Registered: ‎03-15-2019

Re: problem about yolov3 on ZCU104 with dnndk

I din't change anything except the number of classes.
0 Kudos
Xilinx Employee
Xilinx Employee
170 Views
Registered: ‎03-21-2008

Re: problem about yolov3 on ZCU104 with dnndk

So if I undertand correctly, you used the same decent_q and dnnc settings for both the pretrained COCO weights, and you own Yolov3 model.

What about the calibration images, did you use something appropriate for your data set?

0 Kudos
Visitor avisakio
Visitor
137 Views
Registered: ‎03-15-2019

Re: problem about yolov3 on ZCU104 with dnndk

I used a subset of my own training data as the calibration images
0 Kudos