MAIX Kendryte 210 guide

Does any one know if there is a guide to use edge with a the cheap maix boards, I have one but the guides are gritty, reading chinese isn’t my strong point.

If you export as C++ library and you have a way to compile for this target the resulting model will probably run as-is, but will fall back to software implementation for DSP and Neural Network code so it’s too be seen how fast it is. For data collection you can use the data forwarder (https://docs.edgeimpulse.com/docs/cli-data-forwarder) - if your data is not too high resolution.

What are you planning to do with it?

Actually reading here (https://www.hackster.io/dmitrywat/object-detection-with-sipeed-maix-boards-kendryte-k210-421d55) it seems that they use (copied?) the OpenMV IDE and seem compatible with plain tflite formats, so for anything image related perhaps exporting as OpenMV library works.

I want to do hand sign and simple word recognition, the devices come with a mic and a camera.

Collecting data with edge seems way easier than trying to do it in an ad hoc manner.

Yep, should be fine I think - just try and see how fast it is. Hand sign should be fine to grab data with the data forwarder, and software implementation in the C++ library export should be fast enough.

For word recognition it’s a bit harder to collect data from the device, but you can always use your phone for data collection, or use a prebuilt dataset. Then deploy to device in the same manner, and would be interested to see if it runs fast enough!

@marian How did it go for you? Were you able to use EI with K210? I am planning to do the same.

Going the route of compiling standalone C++ library might work, but it will only use K210 CPU and not the neural network inference accelerator(KPU), which means it’s going to be quite slow.
If you want to use EI, you can train your model in EI and then download the .tfilte and convert it with nncase to .kmodel, native model format for K210.

5 Likes

Thanks. Will try this.

@AIWintermuteAI @mithundotdas did either of you have luck with this?

I’ve tried the nncase conversion via docker, but have hit a number of snags, including outdated docs and a seg fault.

If you’re encountering problems with nncase usage or running converted model on K210, I advise you to contact nncase/K210 software maintainers :slight_smile:
I just want to point that if you’re trying to run the converted model (.kmodel) with MaixPy, you can only use kmodel v4, which can be obtained with nncase v0.2.0-beta4 https://github.com/kendryte/nncase/releases/tag/v0.2.0-beta4. Last time I tried classifier conversion it worked without any issues.

1 Like