# NaNsense

> PyTorch debugger: pause training, step batch by batch, and look inside every layer — activations, gradients, weights, optimizer state.

NaNsense is a PyTorch debugger that works inside the live training loop: it pauses training so you can step batch by batch, time-travel to earlier epochs, and inspect the activations, gradients, weights and optimizer state of every layer, plus run interpretability experiments (deep dream, Grad-CAM and other attributions) on the paused model. Each page below also exists as plain Markdown at the listed .md URL. For library internals, see https://github.com/kongaskristjan/nansense/blob/main/INTERNALS.md

## Usage

- [Home](https://kongaskristjan.github.io/nansense/dev/index.md)
- [Showcase](https://kongaskristjan.github.io/nansense/dev/showcase/index.md)
- [Getting started](https://kongaskristjan.github.io/nansense/dev/getting-started/index.md)
- [Integrate with one prompt](https://kongaskristjan.github.io/nansense/dev/integrate/index.md)
- [Wiring guide](https://kongaskristjan.github.io/nansense/dev/wiring/index.md)
- [UI guide](https://kongaskristjan.github.io/nansense/dev/ui/index.md)

## API

- [API reference](https://kongaskristjan.github.io/nansense/dev/api/index.md)