read_tensorboard_scalars

read_tensorboard_scalars(
    logdir
    *
    tags=None
    runs=None
    smoothing=DEFAULT_SMOOTHING
    max_points=DEFAULT_MAX_POINTS
)

Read the scalar charts of a TensorBoard log directory.

One chart per tag, as TensorBoard’s Scalars dashboard draws it: the step along the x axis, the value along the y axis, and one line per run.

Parameters

Name	Type	Description	Default
logdir	str or os.PathLike	The directory TensorBoard would be started with (--logdir), such as the log_dir given to Keras’s TensorBoard callback.	required
tags	iterable of str	Only these tags, in this order. By default every scalar tag, sorted.	None
runs	iterable of str	Only these runs, such as ["train", "validation"].	None
smoothing	float	TensorBoard’s smoothing weight, from 0 up to but not including 1. Above 0, each run is drawn twice: as logged, and smoothed, under the run’s name followed by (smoothed), as TensorBoard shows both. 0 draws the values as logged only.	0.6
max_points	int or None	At most this many points per line, evenly spaced over the steps and always keeping the first and last, so a long run stays quick to walk. None keeps every point.	1000

Returns

Name	Type	Description
	list of TensorBoardChart	One per tag. Pass one to :func:maidr.show, :func:maidr.render or :func:maidr.save_html.

Raises

Name	Type	Description
	FileNotFoundError	If logdir is not a directory.
	ValueError	If smoothing is not in [0, 1) or max_points is below 2.

Notes

Read: scalars written by Keras’s TensorBoard callback, tf.summary.scalar in TensorFlow 1 and 2, PyTorch’s torch.utils.tensorboard.SummaryWriter and tensorboardX. Histograms are read by :func:read_tensorboard_histograms and a Keras model’s graph by :func:read_tensorboard_graph; distributions, images and the other dashboards are not read yet.

Smoothing is TensorBoard’s: an exponential moving average, corrected so the first values are not pulled toward zero. It is computed over every value logged, before max_points thins the line. A value that is not finite, such as a loss that became NaN, is a gap in the line and is left out of the average, as TensorBoard leaves it out.

A damaged event file is read up to the damage, with a warning, wherever the damage reaches a record’s length. Damage inside a record’s data is not detected, since checking it would mean a checksum over every image and histogram the log directory holds, and reads as the damaged value.

Examples

>>> import maidr
>>> charts = maidr.read_tensorboard_scalars("logs/fit", tags=["epoch_loss"])
>>> maidr.save_html(charts[0], "loss.html")