read_tensorboard_projector

read_tensorboard_projector(
    logdir
    *
    label=None
    coordinates=None
    max_points=DEFAULT_MAX_POINTS
)

Read the embeddings of a Projector log directory as 2D scatter plots.

Parameters

Name	Type	Description	Default
logdir	str or os.PathLike	The directory holding `projector_config.pbtxt`, such as the `log_dir` of PyTorch’s `SummaryWriter.add_embedding`.	required
label	str	The metadata column the points are grouped by. By default `label` when there is one, else the first column.	`None`
coordinates	numpy.ndarray	Two coordinates per point to draw instead of the principal components, such as a t-SNE or UMAP projection computed elsewhere. Only for a directory holding one embedding.	`None`
max_points	int or None	At most this many points per chart, evenly sampled, with a warning.	`2000`

Returns

Name	Type	Description
	list of TensorBoardChart	One per embedding, tagged with its `tensor_name`; `runs` names the label groups, in the order their layers are drawn.

Raises

Name	Type	Description
	FileNotFoundError	If `logdir` holds no `projector_config.pbtxt`.
	ValueError	If `coordinates` is not two per point, or the directory holds more than one embedding for it.

Notes

Each label group is one scatter layer, named in the legend, so a reader moves between classes with Page Up and Page Down and along one class’s points with the arrow keys. The axes are the first two principal components, each named with the share of the variance it explains.

Examples

>>> import maidr
>>> (chart,) = maidr.read_tensorboard_projector("runs/embeddings")
>>> maidr.show(chart)