Is it possible to download tensorflow model to FlexRIO+1483 or 1477?

[jiangliang]

Hi:

I made a simple denoising autoencoder DL model, it works at PC(tensorflow with keras), but I'd like to have it gives result in real-time, I noticed FlexRIO have 1000+ DSP48E unit, it should be able to do simple Deep-learning work.

My model saved as h5 file is no more then 10Mbyte, which is really small, my goal is put this model to run in hardware level.

I have both FlexRIO 7935+1483, and I got a 1477 cameralink card recently, I think both device have enough power to run my model at reasonable speed.

(If a raspberry-pi can do this job, why can't the mighty FPGA do it?)

[SamiF]

Maybe you could ask these guys : https://www.ngene.co/dnn-accelerator-for-fpga

Sami

[jiangliang]

Thank you SamiF!

These guys are really good, I did wished NI can offer such toolkits, but it comes out 3rd party is doing a even greater job, I was always hope to use Labview to train NN, because it would be much faster if we can train while acquire real-world data at the same time.

But come back to my problem, they did say they can use FPGA, but it comes out that this toolkit will set the whole FPGA just like GPU, it comes with a FPGA bitfile, the whole FPGA will not at my control if I want to use NN function.

I was hoping I can set part of FPGA for Cameralink I/F, and do some real-time process first, then I'll send the data to NN and fetch the outcome at PC.

[SamiF]

Understood.

Thank you for the information and sorry if you still don't have a solution.

Sami