HiddenLayer is a lightweight Python library for visualizing neural networks and their training dynamics in PyTorch, TensorFlow and Keras. It renders readable architecture graphs and live training-metric plots inside Jupyter notebooks, offering a simpler alternative to heavyweight dashboards for everyday model work. It is an open-source project on GitHub with around 1,900 stars, last updated in 2024. (Note: this is the open-source visualization library by Waleed Abdulla and Phil Ferriere, unrelated to the AI security company of the same name.)
The library traces a model’s computation graph and draws it as a clean diagram you can save to PDF or PNG, which is invaluable for papers, code reviews and debugging unexpected layer wiring. During training, its canvas API plots losses and metrics as they evolve, with support for combining multiple runs for comparison. HiddenLayer deliberately keeps its API tiny โ a few lines turn a Keras, PyTorch or TensorFlow model into a graph โ and it is designed to be extended with custom graph transformations and themes.
Free. HiddenLayer is open source under the MIT license.
The project grew out of the authors’ frustrations explaining model architectures in papers and blog posts, and its example notebooks for each supported framework double as a quick tour of the API. Graph rendering is built on Graphviz, and transformation hooks let users collapse or annotate parts of a network โ for instance folding batch normalization โ so diagrams stay readable even for deeper models. Despite its age, it remains a staple recommendation whenever someone asks how to draw a neural network from code.
Students, researchers and practitioners who want quick architecture diagrams and training curves in notebooks without adopting a full MLOps tracking stack.