modelvision 0.1.0

Framework-agnostic neural network architecture visualizer — publication-quality diagrams without executing a forward pass.

ModelVision

Framework-agnostic neural network architecture visualizer. Renders publication-quality, fully styleable diagrams for models built in PyTorch, TensorFlow/Keras, JAX/Flax, JAX/Haiku, Hugging Face Transformers, scikit-learn, and ONNX — without running a forward pass.

Status: early development (Week 0 foundations landing; M1 next). See PRD_ModelVision.md for the full spec.

Install

uv add modelvision                    # core only
uv add "modelvision[torch]"           # per-framework extras
uv add "modelvision[all]"             # everything

Quick start

import torchvision.models as models
import modelvision as mvision

model = models.vgg16()
mvision.render(
    model,
    output="vgg16.svg",
    theme="dark",
    layer_palette={
        "Conv2d":    "#4a90d9",
        "ReLU":      "#27ae60",
        "MaxPool2d": "#e67e22",
        "Linear":    "#9b59b6",
    },
)

CLI

uvx modelvision model.py MyNet --output diagram.svg --theme dark
uvx modelvision model.onnx --output diagram.html

Development

git clone https://github.com/pianistprogrammer/ModelVision.git
cd modelvision
uv sync --all-extras --extra dev
uv run pytest

License

MIT