Actions
Forum Thread Details
  • Replies 27 replies
  • Subscribers 2558 subscribers
  • Views 3156 views
  • Users 0 members are here
Related

Seeking Your Feedback on Roadtesting an Artificial Intelligence Eval kit

rscasny

imageI've been speaking with a sponsor about roadtesting its new AI Eval kit for machine learning applications.

I'd like to get your interest in this kit, so I've posted below some info about it.

The sponsor was interested in learning how you would roadtest this kit. They would like to see your ideas to better define the goals of this roadtest. So, if you have any thoughts feel free to add any comments about roadtesting below. Thanks.

What is the kit?

Infineon’s PSoCTm 6 Artificial Intelligence Evaluation Kit is a hardware platform focused on Edge AI. It enables customers to evaluate Infineon’s Machine Learning platform IMAGIMOB Studio, as well as ready to deploy ML models, and other SW products.

Here are some more info and features:

PSoCTm 6 Artificial Intelligence Evaluation Kit (CY8CKIT-062S2-AI)

The PSoCTm 6 AI Evaluation Kit is a Hardware platform focused on Machine Learning (ML) . It enables customers to evaluate Infineon’s ML platform Imagimob Studio, as well as ready to deploy ML models, and other software products. Data collection is enabled via PSoCTm 6 MCU, radar, acoustic, pressure & IMU sensors. The board is designed for easy prototyping and lets customers collect real-life data to build compelling ML products fast.

Summary of Features

  • Small form factor
  • Wireless
  • Full ML to embedded Software journey
  • Powerful AI toolchain
  • Production ready ML models
  • Multi sensor input

Potential Applications

  • Sound detection (siren, cough, snore, baby cry)
  • Presence detection, counting & tracking
  • Gesture sensing
  • Customized radar
  • Motion detection
  • Sensor fusion

Here is some documentation:

Datasheet

Schematics

Release Notes

Imagimob AI 

ModusToolBox 

Other Information 

Parents

  • I've played with some AI-focused boards in the past, and so far one thing that is *very* important to test, is what the workflow looks like when it comes to training models. I remember one specific board where all the training had to be done on a proprietary cloud platform where you had to upload the training data, set inputs and outputs, and it would do the training for you. It didn't have a lot of options and the whole process was a black box. The output was an obfuscated binary that had to be uploaded to the board. Training and testing models was a truly miserable experience. And I could never see the whole thing working because the board died in one of the many steps of the process.

    On the other hand I also played with a board that had some level of compatibility with pytorch models, and that made it a lot easier to do all the tinkering, training and testing on my PC, and then just upload the trained model to the board.

    So the "training new models" workflow is what I'd love to test.
    That, and performance.

    EDIT: I see that the toolkit and platform for AI training is mentioned in the description, btw, and I know most of that toolchain is designed to make things easier for users, but I'd be interesting in checking whether the experience is closer to the first example I mentioned, or the second. And whether it's possible to do more traditional learning and then connect that somehow to their toolchain.

Reply

  • I've played with some AI-focused boards in the past, and so far one thing that is *very* important to test, is what the workflow looks like when it comes to training models. I remember one specific board where all the training had to be done on a proprietary cloud platform where you had to upload the training data, set inputs and outputs, and it would do the training for you. It didn't have a lot of options and the whole process was a black box. The output was an obfuscated binary that had to be uploaded to the board. Training and testing models was a truly miserable experience. And I could never see the whole thing working because the board died in one of the many steps of the process.

    On the other hand I also played with a board that had some level of compatibility with pytorch models, and that made it a lot easier to do all the tinkering, training and testing on my PC, and then just upload the trained model to the board.

    So the "training new models" workflow is what I'd love to test.
    That, and performance.

    EDIT: I see that the toolkit and platform for AI training is mentioned in the description, btw, and I know most of that toolchain is designed to make things easier for users, but I'd be interesting in checking whether the experience is closer to the first example I mentioned, or the second. And whether it's possible to do more traditional learning and then connect that somehow to their toolchain.

Children
No Data