torchui

See inside your PyTorch model, in the browser, with one line of code.

import torchui
torchui.watch(model)

torchui is a learning-oriented visualizer for PyTorch. It is not a production experiment tracker. It exists to make the fundamentals visible: what a network actually is as a set of matrices, how data flows through it, what the autograd computation graph looks like, and what really changes in the weights during training.

What it will show

• Architecture as a layer diagram with real tensor shapes flowing through it.
• Weights as matrices: heatmaps of every parameter, conv kernels as images, and a scrubber to watch weights drift from their initial values step by step.
• The computation graph, taken from the real autograd graph (grad_fn), so what you see is what backprop traverses.
• Training dynamics: loss, per-layer gradient norms, update ratios, activation distributions, and dead-ReLU fractions over time.
• Single-sample traces: pick one input and see every intermediate activation on its way to the output.

How it works

torchui.watch(model) registers forward hooks and gradient hooks on the model, observes backward() and optimizer.step() to find the loss and step boundaries, and records everything to a run directory on disk. A small background server serves a static single-page viewer that reads that directory, so you can watch live or replay any past run with torchui runs/<name>.

Status

Early development. This release reserves the package name; the first working version is on its way.

License

MIT