In the description of the architecture of TensorFlow lite for microcontrollers, I read that the model should fit in memory, but which one ?
I understood that the model is not constrained to fit in RAM (am-I correct ?)
But is the model constrained to fit in flash ?
We have access to sdcard on our esp32 and we would like to know if we can use the Sdcard.
I understand that it will definitely be slow and power hungry but it might be interesting anyway for multi-stage inference:
- 1st model detects event A and is fast (fit in ram
- 2nd model refines from A but needs a bigger model that may not fit in flash. As event A is rare in that case we can afford time and energy.