nn3d

Watch neural networks train in live 3D in your browser. Every layer's activation and every connection's weight stream in real time.

import nn3d

nn3d.show(model)                                       # opens browser
model.fit(X, y, callbacks=[nn3d.Monitor(X[:1])])       # training streams live

In a notebook, nothing else is needed. In a plain .py script the process exits when the script finishes — add nn3d.wait() at the end to keep the server alive.

Screenshots

Every image below is a real capture of the viewer while a Keras model was actually training — note the CANLI · epoch N · adim N badge in the bottom-left corner and the metrics ticking in the top-right.

Hover a neuron — its connections light up in gold, everything else dims, and the tooltip reports the real neuron index behind the drawn subset:

A CNN — Conv2D → BatchNorm → MaxPool → Conv2D → MaxPool → Flatten → Dense → Softmax on Fashion-MNIST. Each card shows the layer's tensor shape (26x26x32), and the 784 and 1600 layers are drawn as 16 representative neurons each:

Drag to orbit — the scene is real 3D, so the layer planes separate in depth as you rotate:

Installation

pip install nn3d

Node.js is not required — the 3D viewer ships pre-built inside the package. Node is only needed if you want to modify the viewer itself.

What you see

• Layer cards — name, neuron/channel count, tensor shape, activation like [ReLU]
• Neurons — brightness and size reflect the current activation value
• Connections — teal = positive weight, red = negative; brightness = |weight| × source activation
• Input/output labels — feature names ("Hunger Level", "Fox Distance") color-shift with activation
• Hover — hovering a neuron highlights all its connections in gold, dims the rest
• Metrics — loss, accuracy, val_loss… live in the top-right corner

API

nn3d.show(model, ...)

Builds the graph and opens a browser tab. Returns a View.

nn3d.Monitor(sample, every=20, ...)

A Keras callback. Drop it into model.fit(callbacks=[...]).

every sends a frame every N batches. Running an extra forward pass every single batch adds measurable overhead; every 20 looks smooth and costs almost nothing.

View.update(x, metrics=..., epoch=...)

For custom training loops (RL, manual gradient steps) — push a frame manually.

Virtualization — is what you see real?

Yes, but not all of it. Drawing Dense(3072) as 3 072 points means 2.3 million edges to the previous layer — the browser would freeze.

Instead, max_neurons evenly-spaced representative neurons are selected per layer; activations and weights are sliced to the same indices. Every point and every line you see corresponds to a real neuron or weight — nothing is made up, it is just a subset. When you hover a neuron the tooltip shows the real neuron index (neuron 1847 / 3072) so you always know which slice you are looking at.

Supported layers (Keras 3)

Dense · Conv1D/2D/3D · Conv2DTranspose · SeparableConv2D · LSTM · GRU · SimpleRNN · Flatten · Reshape · Dropout · BatchNormalization · LayerNormalization · MaxPooling* · AveragePooling* · Activation · LeakyReLU · ReLU · Softmax · Embedding

Weights are reduced to a (input_units, output_units) matrix for visualization: conv kernels are summed over spatial axes, LSTM/GRU gates are averaged. The goal is sign and magnitude of the connection, not numerical precision.

Architecture

Keras model ──build_graph()──▶ graph dict ──┐
                                             ├─▶ stdlib HTTP + SSE ──▶ Three.js viewer
ActivationTap.read(x) ──────▶ frame (base64 f32) ┘      (:8092)         (browser)

• Zero dependencies. The server is pure Python standard library. SSE is used instead of WebSockets because the stream is one-way (Python → browser), SSE is exactly what it was designed for, works with http.server, and reconnects automatically on drop. WebSockets would need either an external package or a hand-written handshake — both break the "pip install nn3d is enough" goal.
• The viewer knows nothing about the model — only the schema in src/nn3d/schema.py. The same viewer renders a 10-neuron game agent and a 12-block Transformer. Adding support for another framework (PyTorch, Unity, JS game) means writing a new adapter, not touching the viewer.
• Single geometry. All edges live in one LineSegments, all neurons in one Points. Only the color buffer is updated per frame; geometry never changes.

Modifying the viewer

cd viewer
npm install
npm run dev      # live dev server (start the Python server separately first)
npm run build    # compiles into src/nn3d/static/index.html

The compiled output is part of the package, so the result of npm run build must be committed too.

Examples

PYTHONPATH=src python examples/churn_canli.py