A simple and compact one-line progress bar for TensorFlow 2 Keras.

---

TensorFlow Compact Progress Bar

Existing ways of monitoring training of TensorFlow Keras models either display nothing (verbose=0), one progress bar per epoch (verbose=1), or prints one line of metrics per epoch (verbose=2). When training for thousands of epochs, this often leads to bloated log files or crashed/sluggish Jupyter environments when working interactively.

This library provides a compact progress bar that simply displays the overall training progress by epoch. There are also a few small additional features for convenience, such as excluding certain metrics to avoid excessive clutter.

Notebook mode

Console mode

Getting Started

Install from PyPi:

$ pip install tf-compactprogbar

Dependencies

• tensorflow >= 2 (TF 1 likely works, but untested)
• tqdm
• python >= 3.7
• ipywidgets (Optional)
• jupyterlab_widgets (Optional)

For nice looking Jupyter/IPython progress bars, make sure you have ipywidgets and jupyterlab_widgets if you are on Jupyter Lab.

Usage

To use it, disable the built-in logging (verbose=0) and pass it in as a Callback:

from tf_compactprogbar import CompactProgressBar

progBar = CompactProgressBar()
history = model.fit(X_train, Y_train,
                    epochs=200,
                    batch_size=100,
                    verbose=0,
                    validation_data = (X_test, Y_test),
                    callbacks=[progBar])

Documentation

# Call signature
CompactProgressBar(show_best=True, best_as_max=[], exclude=[], notebook='auto', epochs=None)

Args:
        - show_best    (bool)      Display best metrics. Default: True
        - best_as_max  (list)      Metrics which should be maximized (see note)
        - exclude      (list)      Metrics which should be excluded from display
        - notebook     (str/bool)  Whether to use IPython/Jupyter widget or console. Default: 'auto'
        - epochs       (int)       Optional total number of epochs. Default is inferred from `.fit`.
        
Note: When using `show_best`, by default the "best" metric is the minimum. Pass
in the metric name to `best_as_max` to change this behavior.

If there are any issues in Jupyter, please see the tqdm Issues page for help or disable notebook mode with notebook=False.