What it does HiddenLayer generates clean architecture diagrams of neural networks and plots training metrics directly inside Jupyter Notebook. It supports PyTorch, TensorFlow, and Keras, and can export to PNG or PDF when you’re running headless on a server.

The interesting bit The “hidden” in HiddenLayer is the clever part: it deliberately collapses low-level noise. Conv→ReLU→MaxPool sequences fuse into single boxes, and you can write custom graph-expression rules—like folding an entire ResNet bottleneck block into one labeled node. The authors explicitly position it as “too big for the task” insurance against TensorBoard.

Key highlights

• Renders high-level graphs that hide weight initializations, gradients, and internal ops by default
• Custom folding via ht.Fold() graph expressions (the ResNet example is genuinely neat)
• Live training metrics with extensible Canvas class—subclass it to add pie charts, confusion matrices, whatever Matplotlib can draw
• Works outside Jupyter: pops a window in CLI scripts, or saves PNG snapshots for remote servers
• ~1.8k stars, MIT licensed, Python 3.6 only per the README

Caveats

• Only tested on Python 3.6; older versions unsupported
• Authors warn it’s an “early release” and bugs are likely
• No optimization for large-scale experiment logging yet

Verdict Grab this if you want quick, hackable visualizations without configuring a TensorBoard server. Skip it if you need multi-run comparison, hyperparameter search visualization, or production monitoring—those remain TensorBoard territory.