Working on a U-Net image segmentation project that will use the OpenMV Cam H7 Plus.
I’ve trained a model using Tensorflow 2.8, fully quantized to a .tflite file using int8. Using netron, it seems that the layers should all be supported (referencing this for supported ops tensorflow/all_ops_resolver.cc at master · openmv/tensorflow · GitHub)
Searching around the forums I thought the main reason this would happen is because of unsupported layers, but I’m pretty sure I’ve resolved this.
I have also tried models quantized for float32 and uint8, but the same rebooting behavior occurs.
Could it be because I am training a model using Tensorflow 2.8 and not an older version? What version of Tensorflow should be used to train a custom model?
I would use Edge Impulse if they had features for the model I am trying to make. My model is based on U-Net, so I am getting per-pixel inference which is different from EI’s FOMO implementation.