An experiment: reading digits

The example examples/g_experiments/g3_mnist.py trains a small neural network on handwritten digits, live, in the editor. It uses a picture on a card, a job that reports as it works, a Plot, and Run automatically, so it shows what the pieces do together.

Run it from the repository root, after building the browser layer:

uv run examples/g_experiments/g3_mnist.py

The first start downloads the data: MNIST (60,000 training pictures of digits and 10,000 test pictures) and Fashion-MNIST (the same sizes, pictures of clothes), 42.5 MB in all, into __data__/ in the folder you run it from. Each file is checked against a sha256 pinned in the example before it is kept.

The three nodes

MnistSample has an index box and a dataset box that holds mnist or fashion. It outputs the picture (shown on its card), its label and the dataset word.

Train is a job node. Its boxes are hidden (the size of the hidden layer, 128), learning_rate (0.1), epochs (30) and dataset, which is linked to the sample’s so the two agree. It learns by plain minibatch gradient descent on batches of 64 pictures, written in numpy. Every 100 batches it reports one point of the mean loss and one of the accuracy on the first 1,000 test pictures; its last point uses all 10,000.

A Plot node draws both series as they arrive.

Saving the curves

After a run, the card menu of the Plot (the three bars) has Save as CSV. The file is named for the node, n3-chart.csv, and holds an index column, then loss and accuracy named as the chart’s legend names them. There is one row per point; read it with numpy.loadtxt(path, delimiter=",", skiprows=1) or pandas. A gap, or a line that ends early, is an empty cell.

What to expect

On the default settings MNIST ends at about 97.7% test accuracy in about 7 seconds, and Fashion-MNIST at about 87.8%. A network with 16 hidden units ends at about 95.4% in 3 seconds. These are measured on one Mac with numpy, and a run is the same every time because the random generator starts from a fixed seed.

The example does not use a GPU, a convolution or a saved model. Its network is the smallest one that reads digits well, so you can see every number it uses.