livelossplot

Don't train deep learning models blindfolded! Be impatient and look at each epoch of your training!

🎉 New release 0.5.0 (20 Mar 2020). See CHANGELOG and updated EXAMPLES IN COLAB. There are some API changes, to make it better, cleaner, and more modular. From 0.5.1 (28 May 2020) there is Bokeh support. 🎉

A live training loss plot in Jupyter Notebook for Keras, PyTorch and other frameworks. An open-source Python package by Piotr Migdał, Bartłomiej Olechno and others. Open for collaboration! (Some tasks are as simple as writing code docstrings, so - no excuses! :))

from livelossplot import PlotLossesKeras

model.fit(X_train, Y_train,
          epochs=10,
          validation_data=(X_test, Y_test),
          callbacks=[PlotLossesKeras()],
          verbose=0)

(The most FA)Q: Why not TensorBoard?
A: Jupyter Notebook compatibility (for exploration and teaching). The simplicity of use.

Installation

To install this version from PyPI, type:

pip install livelossplot

To get the newest one from this repo (note that we are in the alpha stage, so there may be frequent updates), type:

pip install git+git://github.com/stared/livelossplot.git

Examples

Look at notebook files with full working examples:

keras.ipynb - a Keras callback
minimal.ipynb - a bare API, to use anywhere
bokeh.ipynb - a bare API, plots with Bokeh (open it in Colab to see the plots)
pytorch.ipynb - a bare API, as applied to PyTorch
2d_prediction_maps.ipynb - example of custom plots - 2d prediction maps (0.4.1+)
poutyne.ipynb - a Poutyne callback (Poutyne is a Keras-like framework for PyTorch)
torchbearer.ipynb - an example using the built in functionality from torchbearer (torchbearer is a model fitting library for PyTorch)
neptune.py and neptune.ipynb - a Neptune.AI

You run examples in Colab.

Overview

Text logs are easy, but it's easy to miss the most crucial information: is it learning, doing nothing or overfitting? Visual feedback allows us to keep track of the training process. Now there is one for Jupyter.

If you want to get serious - use TensorBoard, . But what if you just want to train a small model in Jupyter Notebook? Here is a way to do so, using livelossplot as a plug&play component

from livelossplot import ...

PlotLosses for a generic API.

plotlosses = PlotLosses()
plotlosses.update({'acc': 0.7, 'val_acc': 0.4, 'loss': 0.9, 'val_loss': 1.1})
plot.send()  # draw, update logs, etc

There are callbacks for common libraries and frameworks: PlotLossesKeras, PlotLossesKerasTF, PlotLossesPoutyne, PlotLossesIgnite.

Feel invited to write, and contribute, your adapter. If you want to use a bare logger, there is MainLogger.

from livelossplot.outputs import ...

Plots: MatplotlibPlot, BokehPlot.

Loggers: ExtremaPrinter (to standard output), TensorboardLogger, TensorboardTFLogger, NeptuneLogger.

To use them, initialize PlotLosses with some outputs:

plotlosses = PlotLosses(outputs=[MatplotlibPlot(), TensorboardLogger()])

There are custom matplotlib plots in livelossplot.outputs.matplotlib_subplots you can pass in MatplotlibPlot arguments.

If you like to plot with Bokeh instead of matplotlib, use

plotlosses = PlotLosses(outputs=[BokehPlot()])

Trivia

It started as this gist. Since it went popular, I decided to rewrite it as a package.

Oh, and I am in general interested in data vis, see Simple diagrams of convoluted neural networks (and overview of deep learning architecture diagrams):

A good diagram is worth a thousand equations — let’s create more of these!

...or my other data vis projects.

Todo

If you want more functionality - open an Issue or even better - prepare a Pull Request.

Eduardishion/livelossplot