Importing from TensorBoard

nuboard import-tensorboard turns existing TensorBoard logs into NuBoard runs, so old experiments sit next to new ones in the same viewer and analysis API.

nuboard import-tensorboard runs/ -p mnist   # the folder you pass to tensorboard --logdir
nuboard up                                  # view them (./experiments by default)

imported  baseline/train: 3000 scalars, 52 media, 4 hparams
imported  baseline/validation: 300 scalars, 0 media
imported  lr3e-4/train: 3000 scalars, 52 media, 4 hparams
3 run(s) imported into experiments/mnist. View them with: nuboard up experiments

It needs neither TensorFlow nor TensorBoard: the event files are decoded with the standard library, so the command works with a plain pip install nuboard.

Options

Option	Default	Meaning
LOGDIR		the TensorBoard log folder, or a single run folder
-p, --project	the name of LOGDIR	project to import into
--dir	$NUBOARD_DIR or ./experiments	root folder for projects, as in nb.init
--tag	tensorboard	tag for the imported runs; repeat it for several

What is imported

TensorBoard	NuBoard
each folder with events.out.tfevents.* files	one run, named after its path under LOGDIR, finished at its last event
the folder's parent path (baseline in baseline/train)	the run's group, so group by group puts train and validation together
scalars (add_scalar, tf.summary.scalar)	scalars, with the original steps and wall times
images	PNG or JPEG images; several images in one summary become tag/0, tag/1, ...
videos (add_video, written as animated GIFs)	videos
audio	audio clips, playable in the waveform player
text	text entries
PR curves (add_pr_curve, pr_curve)	PR curves, with the original counts per threshold
histograms (add_histogram, tf.summary.histogram)	distributions, re-binned into 50 even bins; TF1 and PyTorch histograms also keep their mean and standard deviation
meshes (add_mesh)	point clouds of the vertices, with their colors; faces are left out
hyperparameters (add_hparams, hp.hparams)	the run's config, so they appear in the table, sweeps and queries
graphs, embeddings, other plugins	skipped, and counted in the output

Runs keep their original start and end times, so the Timeline page shows when they actually ran. Each run's notes say which folder it came from.

Writer-specific details:

• TensorFlow 2 / Keras: logs/fit/<run>/train and .../validation become two runs in the group fit/<run>. Integer scalars (such as a step counter logged as int64) are imported as numbers.
• PyTorch SummaryWriter: add_hparams writes to a timestamped sub-folder; it is folded into its run, so the run gets the hyperparameters and the hparam/... scalars. add_scalars("loss", {...}) writes one sub-folder per key, as TensorBoard shows them, and each becomes a run.
• tensorboardX and TensorFlow 1: scalars, images and audio are read from the classic summary fields.

Importing again

Run ids are derived from the source folder and its first event, so importing the same LOGDIR again skips runs that were already imported and adds any new ones. To re-import a run, delete its folder under <project>/runs/ first.

A log that ends in a half-written record (a training job that was killed) is imported up to the last complete event.

From Python

from nuboard.tensorboard import import_logdir

for result in import_logdir("runs/", "mnist", dir="experiments"):
    print(result.name, result.run_id, result.scalars, result.media, dict(result.ignored))

nuboard.tensorboard.read_events(path) yields the decoded events of a single file, if you need to read TensorBoard data without importing it.