Deep Learning FPGA Deployment on Xilinx ZCU102

I would like to Run a Deep Learning Networks on Xilinx ZCU102 board, as shown here: Run a Deep Learning Network on FPGA with Live Camera Input - MATLAB & Simulink (mathworks.com)

But I can't find the model "Deep Learning Pre-Process Hardware Algorithm Target Model" mentioned in the link.

Would it be possible to get an off-line model ?

My matlab version is R2021a.

Thanks in advance.

[janny ring]

Any help is much appreciated!

[Steve Kuznicki]

Janny,

I can provide you with the necessary models and files to run this. This requires the ZCU102 and the Avnet FMC-HDMI-CAM card if you want to actually use Live Video. It also requries a usb mouse and micro-2-standard adaptor (to plug into the ZCU102).

Steve Kuznicki

[janny ring]

I already poessess ZCU102 and ready to purchase the Avnet FMC card.How can I get your mentioned models and so on files ?

Thanks for you reply

[Steve Kuznicki]

Janny, You will need to provide me with your email address for me to send this model and support package (for R2021aUpdate4).

Answer 1

[Kamil Sawicki]

Hello,

@janny ring could you please share some of your experience with me regarding example model that you mentioned. Right now I possess zcu102 and wonder if this is right direction before I purchase FMC HDMI sister card. Did you have to make any aditional purchases, what camera did you use and did you manage to use other neural networks with this example?

Thank you for your help in advance.

[Kamil Sawicki]

Probably @Steve Kuznicki could provide some answers, thank you.

[Steve Kuznicki]

Hi Kamil,

A regular 1080p HDMI "source" and "monitor" will do just fine here. I've used a GoPro and Sony Camera and even an IPad or Laptop to provide the HDMI Input. Currently I use a AIDA HDMI camera. The Monitor can be any with an HDMI input although some odd/cheaper ones may not work but I rarely run into that.

Different networks can surely work with this example. The "Pre-Processing" used here is really implementing the ImageInputLayer (zscore or zerocenter normalization) so your network will have to have the same layer, otherwise you can modify the Simulink model and regenerate the bitstream. The Post-processing is done on the ARM Processor and implements the layers the DLP do not execute. You can see which layers are Hardware vs Software (i.e. not executed in the DLP) when you "compile" the network using the Workflow object. So the Resnet18 and the Yolov2Resnet18 networks have different post-processing and therefore 2 different Simulink models are provided.

[Kamil Sawicki]

Thank you for quick answer,

Cameras with HDMI output are not that common, but I understand why this type of input was chose over USB for example, but probably for now I could just stick with PC as a data source. Could you please tell me what tools are required to edit this example, can everything be done with Simulink and Matlab scripts, or would I have to use Vivado in order to get it going with my custom neural network?

Answer 2

[hao]

你好，请问您有实现该例程吗？matlab提供的支持文档对新手不是很友好。