Interact with a native torch session, restored in the browser
We train a 1D-MNIST torch classifier, export it to ONNX, and cache the result.
The cache lets an expensive notebook run in the browser, where PyTorch has no
emscripten target. The training cell wraps a moutils.onnx.OnnxRuntime in
mo.persistent_cache, and moutils’ cache stub serializes that runtime as its
ONNX bytes. On this page it is a cache hit, so torch, pymde, and mnist1d never
import, yet runtime restores as a working session.
Lasso a region of the PyMDE embedding. Then step through the selected samples.
Each sample runs through the restored runtime via onnxruntime-web in the
browser. The confusion matrix summarizes the model’s predictions for the region.
The frame embeds a static marimo export html-wasm --execute build. The notebook
is embeds/onnx_mnist1d.py, and the export bundles moutils as a wheel.