I would like to implement some basic (or not ) ML features.
I am looking for an ev. board that will allow me to process Voice, Images and sensors (Acc, Gyro).
I am not an expert in ML/DL so the goal would be to port existing code and to evaluate the results.
I am looking for a HW that will fit my requirements (I also need BLE & GPS + Camera).
I would be happy to get some advise what to choose for a stable HW with a good support. Feel free to ask me for more info if necessary.
Thanks a lot.
By the way, I am in the process of creating a new concept so I look for serious people to join this adventure. NLP, image processing,ML/DL are the main areas of this concept.
Feel free to contact me if relevant.
I will start to study this board.
The only drawback I can see right now is that it is not working with a known AI framework like tensorflow.
I believe that it can also be an advantage to use an SDK provided by the vendor since it may be very optimized to this specific HW
I saw the boards proposed by Steve but again, many features were missing.
BTW, I am also opened to add add-on modules to the main board.
Thanks again!
If you don’t really need the “Tiny” part of Tiny ML, you might also consider one of the Raspberry Pi boards. This could be especially attractive if you are just starting out and want to experiment with various sensor data streams. The nice thing about the Pi boards is that they can run the same Python that runs on your laptop, and there are a lot of utility libraries to make things easy. I recently paired up the Raspberry Pi 3B with a Sense Hat to build a simple WIFI-based datalogger for accel/mag/gyro/pressure and humidity. Data rates aren’t horribly high (10-20 sps), but that’s primarily because I used Python throughout for speed of implementation. I know you can add camera and external GPS to the PI if desired. The downside for this platform of course is size and power.
I spent a number of years at a major semiconductor maker building ML-based demos on custom sensor boards with our own (and 3rd party) software. I can tell you categorically that you can spend man-years just developing the drivers and communications code. I spend probably only 10% of my time on the ML, and the rest of the time on everything else necessary to build a working system. So you are right to want to leverage existing libraries as much as possible.
The ST B-L475E-IOT01A actually uses TensorFlow Lite for Microcontrollers if you use it with Edge Impulse. ST also has their own CubeAI framework, which might create some confusion, but you don’t have to use it.
Hello Mike.
Thank you very much for your clarifications. It helps a lot.
Actually I do need the “tiny” aspect of the board.
I know that requesting low consumption, high power, standard SDK and support for audio/camera/GPS/… is very challenging and maybe impossible. I will probably have to renounce to something.
Thanks again
yes, you probably will have to compromise somewhere. TinyML power targets are generally 1-10mW, which is really tight. Just processing a microphone data stream will probably push you above that. I don’t know what camera and GPS power levels look like, but I’m guessing they will be an issue as well.