Skip to content
ARENA-educationPublic

About

A library for plotting training curves live as the model trains

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

liveplot

Live training curves in Jupyter, Colab and the VS Code / Cursor interactive window, with a tqdm bar underneath, at (almost) no cost to the training loop.

docs Open in Colab tests

training loss every step, eval loss and accuracy every 50 steps, on one panel with two y-axes

Install

pip install "liveplot[notebook] @ git+https://github.com/ARENA-education/liveplot.git"

matplotlib, numpy and pillow are the requirements. The notebook extra adds ipython for the live display and tqdm for the progress bar. A notebook already has both, so plain pip install git+https://github.com/ARENA-education/liveplot.git is enough there.

Quickstart

from liveplot import LivePlot

plot = LivePlot(loader, "loss | acc", "lr")   # wrap the loop, like tqdm
plot["acc"].set_ylim(0, 1)                    # configure, like matplotlib
for batch in plot:
    loss, acc = train_step(batch)
    plot.log(loss=loss, acc=acc, lr=lr)       # log, like wandb
  • Wrap the loop. LivePlot(loader) yields the loader's items. Each item moves the x-axis one step and updates the progress bar. plot.log on its own does not move x.
  • Log by name. plot.log(loss=loss) adds a point to the curve loss at the current x. A metric exists from the first time you log it.
  • Lay out with strings. Each string is one panel. Names separated by spaces share a y-axis. Names after | go on a right-hand y-axis. Here loss and acc share a panel with two y-axes, and lr has its own panel. With no strings, every metric shares one panel.
  • Configure with matplotlib's names. plot["acc"] is the y-axis that holds acc.

For nested epoch and batch loops, create the plot once and wrap the inner loop with plot(...). The x-axis continues across epochs:

plot = LivePlot("loss", "acc")
for epoch in range(epochs):
    for batch in plot(loader, desc=f"epoch {epoch}"):
        plot.log(loss=train_step(batch))      # one point per batch
    plot.log(acc=evaluate())                  # one point per epoch
plot.finish()

The plot is drawn by a separate process and shown as a plain image, so it behaves the same in every notebook front end, and interrupting a cell is safe. Everything you log is also kept in plot.data.

Documentation

The documentation is at https://arena-education.github.io/liveplot/:

  • Common questions: is total needed, why wrap the dataloader, two metrics on one axis, a right-hand axis, several panels.
  • Guide: the x-axis, panels, logging at different rates, customising axes, reference lines, images, smoothing, saving and recording.
  • API reference: every class and method, generated from the docstrings.

The pages are built from docs/ and the docstrings in liveplot/liveplot.py. To build them locally:

pip install -e ".[docs]"
mkdocs serve

Examples

examples/demo.py is a cell-by-cell tour of the features. Run it in the VS Code / Cursor interactive window, or open it in Colab: CI converts it to a notebook on the demo branch on every push to main. The GIF above is cell 2 of that file, recorded by the library itself in examples/make_gif.py. examples/dcgan_synthetic.py shows loss curves beside generated images.

Credits

The per-panel label/limit options and the grid-shape rule are adapted from Tyler Lum's live_plotter (MIT); see THIRD_PARTY_LICENSES.md.

About

A library for plotting training curves live as the model trains

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages