How is energy consumption is calculated? about mltk HOT 4 CLOSED

Hi, thank you for your feedback.

The energy is estimated using "traditional" machine learning:
https://github.com/SiliconLabs/mltk/blob/master/mltk/core/tflite_micro/accelerators/mvp/estimator/base_estimator.py#L123

For each layer in the model, the various parameters are extracted/calculated and then provided as in input a an energy prediction ML model.
The ML model then generates a "prediction" on the amount of energy that layer would require for the given parameters.
All of the layers' predictions are then summed together to get the total estimated energy required by the model.

Hope this helps.

from mltk.

Thank you for your quick response.
Is there any reference for this "traditional" machine learning model for the estimation?
I'm curious about how accurate these estimations are.

from mltk.

The energy estimating models are "experimental" and the details for them are not published.
They work best with the CNNs, especially standard model architectures such as MobileNet and ResNet.
We were seeing accuracies between 5-20%.

from mltk.

I understand, thanks again. I guess the estimation accuracy is too low for me to use it as reference.
I think this issue can be closed then.

from mltk.