We started off the day continuing to experiment. While mika prepares the software change, Mag records some more loops and creates a soundscape. Today, Mag has to go to Göteborg for her job meeting on Monday, so we have only the morning to work together.
Mag developed a routine of making video blogs each day. She was first shy of filming herself, but now she is getting used to it. For anyone, it is horrible to hear oneself in videos, talking. How stupid you sound to yourself! (but for others you sound exactly how you sound, and not stupid at all!)
This time, we use the blue tape. From yesterday’s experiments, it felt like less sensor the better control. So we try only 4 sensors this time, wrists and under arm for both sides. Also by observing how Mag plays with the analog style patch on Organelle, she mostly uses the 3rd and 4th knobs (resonance and cut-off) so we concentrate on these two control values this time.
The result seems promising. Mag says it feels much more controllable. The response seems also accurate and without big latency despite the ML in the middle. The training was done only with the wrist movements, but as she moves her arm up and down, it adds extra mix in the value that creates nice effect.
The ml-lib ml.ann object on PD has a bit of bug (or maybe I am using it wrong) that when you use liner activation, it overshoot to nan (not a number) range and crashes the whole patch on Bela. When using sigmoid activation instead this does not happen, but liner activation gives much nicer result. Mika is going to look into it while Mag is away.
In the evening, Ingo makes a great fix for ml-lib. The PD version of the ml.ann object fails to save trained model (.model) file and can only save training data (.data) files when you send the “write” command. Because of this, every time you want to load the previous training session, you have to read the training data and retrain it, which could result in slightly different model. He looked into the source code and found a bug in file writing function and fixed it. Now we are compiling on Bela so we can save the trained model next time we try it.
In the evening, I was trying to cook potato and leek soup and found this can of corn that is good until February 2013 in the cupboard. I was not sure if a can of corn gets bad…. but expired for 7 years feels a bit scary and decided not to use it and put it back. Maybe someone will find it another 7 years later…