Playground

Design a network, train it here

Shape a Fourier KAN and watch its size update. Train it right here in your browser with Python, or export a bundle that runs with PyTorch and the Q.ANT toolkit on your own machine, then bring the results back.

1Design

Target

The benchmark used for every measured run.

Hidden layers1

Width16

Grid (cosines per edge)16

Learning rate

Steps

Training noise

Seed

Backends

CPU float32Always runs. CPU, bfloat16 emulatedRounding only. An approximation of the device, not a hardware result. Q.ANT toolkitRuns on a photonic card if one is present, otherwise on the toolkit's CPU emulator. The result records which. Skipped cleanly if the toolkit is not installed. Remote NPUReserved for a hosted queue. Not available yet.

2Architecture

parameters0

time in browser0

layers0

Train in browser runs NumPy in your browser and scores the result three ways: float32, bfloat16 rounding, and the Q.ANT toolkit emulator's arithmetic, replicated exactly. Nothing is uploaded. The first run downloads Python once, about 15 MB.

Export gives you the PyTorch version. Inside: run.py, notebook.ipynb, the btt package and your config.json. Run pip install -r requirements.txt then python run.py, or open the notebook in Colab.

3Results

Ran the exported bundle? Drop its results.json here or click to choose it. It stays in your browser, nothing is uploaded.

No run yet? : the 60,000 step precision floor run from the October rerun, produced by this exact bundle, with the toolkit emulator column included.