TensorFlow Lite experiments

I had a lab assignment in an embedded systems class that I thought was a good opportunity to check out TF Lite.

Given some sample of voltage, determine the frequency of a sine wave signal.

I started by exploring the data measured by my microcontroller and generating some sine waves with gaussian noise. This let me create thousands of training samples very quick.

I tried out a few different models, and I found that adding more dense layers helped with prediction accuracy, but was very expensive in terms of model size. Ultimately, I went with two dense layers with 32 neurons each.

Looking at some tests runs, we can see that the model performs reasonably well.

To prepare the model for an embedded deployment, I convert it to a Lite model (250kb → 40kb) and quantize it (40kb → 10kb). Even at 4% the size, the new model performs well.