keras-tinygrad browser demo hub

build 2026-10-04 10:56 — the single page for everything testable in a browser. Embedded: two Keras-traced WebGPU bundles + tf.js. The Pyodide engine downloads Pyodide + wheels (~25 MB, first use) — the only network this page ever does.

A Keras training step, traced through the tinygrad backend and exported as WebGPU kernels. Weight buffers update in place: looping step(x, y) is SGD training. Pick a model, a task, the engines to race, and a mode.

model Dense 784→128→10
Dense + Dropout(0.3) (on-device threefry masks)
custom (any layers/dims/optimizer JSON — traced live in the tab; pyodide engine only)

task easy (±1 prototypes, noise 0.35: ~20σ margin, solved in ~5 steps; lr 0.05)
hard (±1/12 prototypes, unit noise: ~1.7σ margin, classes overlap; lr 0.005)

engines keras-tinygrad — prebuilt bundle (traced in Python, embedded)
tf.js webgl (same init & batches; dropout: its own masks)
keras-tinygrad — traced in-tab (Pyodide) (keras 3.15.1 in WASM; needs network)

mode train steps
same-batch probe ×3, lr=0 (dense: identical; dropout: distinct = fresh RNG; bundle + pyodide)