Skip to content

HugoTini/DebugTShow

Repository files navigation

DebugTShow

A VSCode(ium) extension to get instant feedback by plotting PyTorch tensors when selecting them during a debug session. The addon is using torchshow to plot the tensor.

demo.webm

Install

  1. Make sure torchshow is installed in your python project : pip install torchshow.

  2. Download the latest .vsix from the release page.

  3. In VSCode / VSCodium, in the extensions panel menu, click Install from VSIX and select the downloaded extension.

Usage

See video above,

  1. Start a debug session in your PyTorch project and pause it (for e.g. with a breakpoint).

  2. Double click on a tensor or select a valid expression (that resolves to a tensor). The tensor must be in a format that torchshow can recognize.

Notes :

  • If the expression is valid and resolves to a tensor that torchshow is able to plot, it will be displayed. Otherwise, it will silently fail and will not update the image. To display error messages, use the commands DebugTShow: Debug Messages Disable & DebugTShow: Debug Messages Enable. Leave it disabled when not needed.

  • The selected text can be anything, comments included. It can be useful to add temporary debug plot snippets (see video above).

  • To quickly disable / enable the extension, use the commands DebugTShow: Disable & DebugTShow: Enable (or completely disable / uninstall it from the extension panel).

  • The displayed image is saved in _torchshow/. If using git, it is recommended to add this folder to your .gitignore.

License

This uses code from vscode-debug-visualizer. This repo is also under GPL.