Changelog

Each release lists the Google Colab runtime it was tested against (from googlecolab/backend-info).

0.1.4 (unreleased)

tools.layer_shapes(): input and output shapes of any layer or block (all positional inputs, keyword tensors, and every output), from one forward pass in eval mode.
viz.show_cam(..., show_color_bar=True): a 0-1 importance color bar next to the overlay.
install() uses Colab’s preinstalled uv (much faster), falling back to pip. backend="auto" | "uv" | "pip". upgrade=True never upgrades dependencies.

torchinfo, netron, onnx, and onnxscript are now installed with the package (versions capped; verified to change no preinstalled Colab package). onnx stays below 1.23, which would require upgrading Colab’s protobuf. cu.install("tools") is no longer needed.
tools.netron() / export_for_netron() export to ONNX by default, with the graph’s input and output named input and output, and hide torch’s internal treespec warning.

viz.show_images(), show_by_class(), show_cam(), show_feature_maps(): the plot no longer appears twice when the call is the last line of a notebook cell.

Tested against: Colab GPU runtime, Python 3.13.15, torch 2.11.0+cu130, torchvision 0.26.0+cu130, fastai 2.8.12.

env: runtime info, notebook name detection, version check against the tested Colab runtime.
env.load_secrets() / env.get_secret(): API keys from the environment, Colab Secrets, a .env file, or a prompt.
viz.show_images(), viz.show_by_class(), viz.tensor_to_image(): look at datasets, loaders, and batches.
drive: Drive mounting, timestamped run folders, persistent config (data_dir, runs_dir).
tools: torchinfo summary wrapper, Netron viewer inside Colab.
install(): locked installer that refuses to change packages Colab already provides.
models: load_model() for fastai exports/checkpoints and PyTorch models/state dicts.
viz: Grad-CAM, layer activations, feature-map plotting.
Sphinx documentation (user guide and API reference), published to GitHub Pages.