Inspect deep learning models as 3D browser wireframes

What it does

TensorSpace turns pre-trained neural networks into explorable 3D diagrams that run in the browser. Built on Three.js and TensorFlow.js, it loads converted TensorFlow, Keras, or TensorFlow.js models and renders each layer as a WebGL object you can orbit around, inspect, and feed data through to watch intermediate inferences propagate. The API borrows Keras layer names—Conv2d, Dense, Pooling2d—so defining a model topology feels familiar even though you’re building a graphics scene, not training.

The interesting bit

Most architecture diagrams are static; TensorSpace treats the network as a physical machine. A separate preprocessing step extracts intermediate layer outputs so the browser can animate how a handwritten “5” trickles through LeNet’s convolutions and pools without ever touching the original Python training code.

Key highlights

• Live demos include LeNet, AlexNet, ResNet-50, VGG16, and YOLOv2-tiny rendered as interactive 3D scenes
• Keras-like Sequential API for stacking layers in JavaScript and wiring them to pre-trained weights
• Visualizes intermediate data manipulations and final inference, not just static topology
• Runs entirely in the browser after the model is converted

Caveats

• Requires a separate TensorSpace-Converter pass to extract intermediate outputs; raw model files won’t load directly

Verdict

Useful if you need to teach, debug, or demo model internals to people who learn better by walking through a space than reading a summary. If you only care about loss curves and confusion matrices, it won’t replace your notebook.