I have been struggling to make the new sleeve shape prototype function for days… and it turned out I made a “simple” copy paste mistake!!! In the python simulation, I connected 2nd hidden layer input to input layer instead of 1st hidden layer, resulting to calculate all the weights and bias wrong.
it should have been like this..
As I do not have extra pins to use software serial as I am using ATTINY85, I had to switch between the pins and do complicated debugging with FTDI. well, good thing is that now the lights are calibrated very well.
Ingo has kindly made me a new dump function so that I can see all the numbers going through the neurons so I do not make this kind of mistake in future. thanks!
and now, it works smoothly. It has some delay, and I am guessing it is coming from big capacitor (47uF) I added to sensor input pins to smooth the PWM. I will switch to 10uF or smaller to see if I can get smooth enough data and gain some speed. I also have 3 loop averaging on software side.