nn_vis VS pyomyo

In my master Thesis I did some work on visualizing a neural network. It is not trivial to show the weights in a meaningful/understandable way. https://github.com/julrog/nn_vis You need a portion of training data and it's just for fully connected layers currently. The visualization is 3D and in realtime, but it needs some preprocessing, so i'm not sure if it fits your needs at all.

I had some some success on pruning weights with adding batch normalization layer between existing layer, freezing the existing layer and then and retrain the model with the batch normalization layer (training can be much shorter because of way less weights to train). Then using magnitude of the original weights with the weights from the batch normalization, you can prune the original model. You can see an example for fully connected layer in my code: https://github.com/julrog/nn_vis

I just made my project public on GitHub, which seems similar to yours https://github.com/julrog/nn_vis

They make devices that do this sort of thing. This project might be worth looking at if you get your hands on the hardware.

For context on the project, I am trying to control some servos using the Myo Armband. I am using PerlinWarp's PyoMyo library to interact with the armband (Github here: https://github.com/PerlinWarp/pyomyo).

I've made a getting started tutorial here which covers the basics, for finger tracking I've made 2 libraries. I recommend starting with predictor_basic in NeuroGloves which is here. For more info or help join the Discord.

I'm slowly cleaning up and open sourcing the project, starting with the driver, pyomyo which everything builds upon, that's available here, which includes an example of using it to play Breakout.

I'm using EMG for this with the Thalmic Labs Myo, I had to make a new library for this device which I've open sourced here.

Cross platform and multithreaded and works without the Myo SDK. GitHub link