Introduction

When TensorBoard shows only some of your experiments, the problem is usually in the log directory layout rather than in TensorBoard itself. Runs disappear when multiple jobs write to the same folder, when event files are incomplete, or when the command points at the wrong directory level.

Give Each Run Its Own Log Directory

TensorBoard treats subdirectories under the log root as separate runs. If two experiments write into the same path, their event files are merged and the UI can look incomplete or inconsistent.

A safe pattern is to generate a unique folder per run:

1import datetime
2import tensorflow as tf
3
4run_name = datetime.datetime.now().strftime("run-%Y%m%d-%H%M%S")
5log_dir = f"logs/{run_name}"
6writer = tf.summary.create_file_writer(log_dir)
7
8with writer.as_default():
9    for step in range(3):
10        tf.summary.scalar("loss", 1.0 / (step + 1), step=step)
11        writer.flush()

Then start TensorBoard at the parent directory:

tensorboard --logdir logs

If you point it at a single run directory, you will only see that one run.

Check the Directory Structure First

A correct layout usually looks like this:

1logs/
2  run-20260311-090000/
3    events.out.tfevents...
4  run-20260311-091500/
5    events.out.tfevents...

If your structure looks like logs/events.out.tfevents... with no per-run subdirectories, TensorBoard has less information to separate experiments cleanly.

You can inspect what was actually written with:

find logs -maxdepth 2 -type f

That is often faster than guessing from the UI.

Make Sure Data Is Flushed to Disk

Another common issue is that the training job is still buffering summaries. TensorBoard can only read what has actually been written to disk.

When debugging visibility problems, flush explicitly:

with writer.as_default():
    tf.summary.scalar("accuracy", 0.92, step=10)
    writer.flush()

If you are using model.fit with Keras callbacks, check that each run gets a unique log_dir value:

callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir)
model.fit(x_train, y_train, epochs=5, callbacks=[callback])

Reusing the same callback directory across experiments is a common reason older runs seem to vanish.

Verify the TensorBoard Command

Sometimes the issue is simply the command line. These are the usual checks:

• '--logdir should point to the parent folder containing runs'
• the process should have permission to read those files
• you should refresh the page after new runs appear

If you want named groups, --logdir_spec can help:

tensorboard --logdir_spec train:logs/train,eval:logs/eval

That does not fix missing files, but it can make the run list easier to understand.

Watch for Mixed Formats or Corrupt Logs

TensorBoard event files are append-only protobuf records. If a process crashes mid-write, or if several jobs write concurrently into one file tree, some summaries may become unreadable.

This is also why copying partially written logs between machines can produce inconsistent results. If one run never appears, test with a clean run in a fresh directory before blaming the UI.

Common Pitfalls

The most common mistake is logging multiple experiments into the same directory. TensorBoard can read that data, but the runs are no longer clearly separated.

Another issue is launching TensorBoard against the wrong folder level. Point it at logs, not at one specific run unless that is intentional.

The third common problem is failing to flush summaries or stopping a job before event files are fully written.

Summary

• Give every experiment its own subdirectory under a shared log root.
• Start TensorBoard with --logdir set to the parent log directory.
• Flush summary writers when debugging missing runs.
• Avoid reusing the same log path across unrelated experiments.
• Inspect the actual event-file layout on disk before assuming TensorBoard is broken.