Train in your browser

Training lab

Pick a dataset, set a model and its training settings, then train on the full training split. The test metric is measured on the full test split after each epoch.

Train a TypeNN model in this page, on the full training split of a real dataset. Your browser downloads MNIST, Fashion-MNIST, IMDB, Reuters and California housing directly from the Keras dataset storage. CIFAR-10 and CIFAR-100 are mirrored on data.typeshade.dev, because their host allows no cross-origin reads. Each card names the citation and the license.

1. Dataset

2. Model and training

Model

MLP: Flatten, Linear, ReLU, Dropout, Linear

Hidden width

Optimizer

Adam

Learning rate

Batch size

Epochs

Execution tier

Auto (WebGPU, then CPU)

3. Run

Not running.

Training loss per step (smoothed)

Test metric per epoch

Kernel graph

The graph shows how TypeNN runs one read: each box is a tensor, and each group is a set of calls that one dispatch runs on WebGPU. A fused group is one dispatch. A dispatch is one GPU call. How a read runs, on GitHub ↗

Choose a button to draw the graph of a read.

4. Results

No finished runs yet.

5. Try the model

Run a model first. Then you can try it here.

What this runs: the TypeNN build shipped with this site, on the tier you select. Each step is one optimizer update over one batch.

Times on this page are wall-clock times in this browser. A software renderer (SwiftShader) is not a physical GPU benchmark.