Accessing tflite model's parameters from Spresense SDK

[suburban-daredevil]

Hi

1. Is it possible to perform On-Device training using the existing TFLM_RT support on Spresense SDK?

2. Is it possible to access / update the model parameters (weights and biases) of the tflite model using Spresense SDK?

Thanks in advance

[CamilaSouza]

Hey, @suburban-daredevil

1. No, that is not possible. The idea is that you train your model on your machine and do only inference on Spresense.

2. It is possible to access the model parameters, but there is no predefined function to do that. But you could read the binary file of the model and modify it if you know what to do.

[suburban-daredevil]

Hey @CamilaSouza

Thanks for the suggestion!

[Benediction 1 2]

Hey @suburban-daredevil, I’m working on something similar to this or exactly this, there hasn’t been a lot of work that I’ve found that relates to this topic, therefore there’s a lot of research that needs to be done. I’d love to bounce off some ideas with you if you’re still researching this topic, and still want to implement on-device training. Please let me know whether you’re interested. Thank you!

[suburban-daredevil]

Hey @Benediction-1-2,

Yeah I'm currently researching in this topic and would to collaborate. Any ideas to move forward with regard to this?

Thanks