Model Memory Issues on Nicla Vision

Arduino Boards
1

Hi everyone,

I’m currently working on a project to detect driver distraction using the Nicla Vision board. I’ve trained a model using MobileNetV2 (96x96x1.0). However, I’m encountering issues when trying to run the .tflite model on the Nicla Vision.

When I attempt to deploy the model, the board crashes and I receive a message indicating that the board detected read storage issues. I can temporarily solve this problem by rebooting the board, but I’m still unable to get the model running successfully.

I’m suspecting that the issue might be related to the model not fitting into the available RAM on the Nicla Vision. Has anyone else experienced similar problems or have any suggestions on how to resolve this? Should I consider compressing the model further, or are there other optimizations I should look into?

Any insights or recommendations would be greatly appreciated!

Thanks in advance.

2

You generally need to compile the model into the firmware for the Nicla. It has limited RAM so the model needs to be baked into flash. EdgeImpulse can do this for you or you can follow the guide here:

openmv/src/lib/libtf at master · openmv/openmv (github.com)

3

I’m not sure if I’m making the right choices, but when I try to build the model for Nicla Vision or Arduino, I always get a failed job.

image
image834×682 22.7 KB

4

image
image615×720 105 KB

5

image
image595×624 25.1 KB

Sorry for the multiple replies; I can’t post multiple images in the same post.

6

image
image1634×980 117 KB

7

It’s because the model is too big.

See what flash text overflowed by. It’s like 2,638,184 bytes… So, you’re 2MB over the limit. You need to re-think the architecture.

You can use that model on the RT1062 or H7 Plus though as they have SDRAM.