the nerwork is a neural network visualisation tool. It's quite slow on scratch, consider turbowarp: https://turbowarp.org/865148967?hqpen keybinds: mouse: draw your own samples (click and drag on layer 1 node square (28 * 28)) space: load random sample (from supply training data) c: clear layer 1 nodes (for drawing your own) n: center drawing t: begin/end training (you may need to hold this down) s: save weights and biases to file (access through saveFile) a: save drawn data to dataset file (access through trainingFile) l: load sample manually (labelled by number) r: randomise weights and biases (reset before training) arrow keys: transform loaded instance
The way to read the nerwork is that the 28x28 body of nodes on the left is the inputs, then there are two hidden layers of 16 nodes each (configurable) and an output layer with 10 nodes (from top to bottom goes from 0 to 9). So when you load or draw a sample, analyse the output layer to see what the nerwork thinks your drawing is. Because it's quite dependent on absolute space (transformations will change the result a lot) it is helpful to hit n to center your drawing for best results This is an scratch adaptation of the c version which can be found here: https://github.com/Known4225/nerwork-OpenGL The first 100 samples in the mnist_train database are preloaded as well as a decent set of weights and biases.
