Everything K-culture — comebacks to K-beauty, straight to your inboxGet it in your inbox

METAL MEDIA

google-ai-edge/model-explorer

1,532JavaScriptApache-2.0

A tool that lets you open up a deep learning model's internal structure as a diagram

Model Explorer visualizes a deep learning model's computation graph as a nested, hierarchical layout. You can expand or collapse layers, highlight input and output operations, and overlay metadata directly on nodes. It supports TFLite, TF, TFJS, MLIR, and PyTorch (Exported Program) formats, with an extension framework for adding more.

What it does

  1. Organizes model operations into nested layers you can expand or collapse to explore the graph
  2. Offers input/output operation highlighting, metadata overlays on nodes, interactive layer pop-ups, search, identical-layer detection, and GPU-accelerated rendering
  3. Installable via pip install ai-edge-model-explorer and launchable with the model-explorer command, or tried directly on a Hugging Face demo with uploaded models
  4. The core visualizer component is also distributed as the ai-edge-model-explorer-visualizer npm package
  5. Community-built adapters exist for ONNX, VGF, and TOSA formats as separate repositories

Why it matters

As models grow more complex, it becomes hard to trace how operations connect just by reading code, and this tool lets you inspect that structure visually while debugging. Its adapter extension framework means support for formats beyond TFLite or PyTorch can be added by anyone.

Terms in this repo

  • TFLite · A lightweight TensorFlow model format for running models on mobile and embedded devices
  • MLIR · An intermediate representation format shared across multiple compilers
  • Exported Program · A graph-form representation of a PyTorch model after being exported
  • Adapter · An extension module that converts a specific model format so the tool can read it

Repository description (English)

A modern model graph visualizer and debugger

Owner · google-deepmind

Open on GitHub

Trending repos

All repos →

Latest from METAL MEDIA