How to log to TensorBoard in Keras

TensorBoard logs let a Keras training run record metrics, graphs, histograms, and profiling data for later inspection. The built-in keras.callbacks.TensorBoard callback writes event files while model.fit() runs.

The callback requires TensorFlow because TensorBoard event writing is TensorFlow-based. Set the log directory per run so event files from different experiments do not overwrite or mix with each other.

Start with epoch-level scalar logging for loss and metrics. Add histograms, images, or profiling only when needed because frequent logging can slow training.

1. Install Keras, TensorFlow, and TensorBoard.

   $ python -m pip install --upgrade keras tensorflow tensorboard

2. Create log_tensorboard.py.

   log_tensorboard.py

   import os
   import shutil
   from pathlib import Path
    
   os.environ["KERAS_BACKEND"] = "tensorflow"
   os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"
    
   import keras
   import numpy as np
    
    
   keras.utils.set_random_seed(31)
    
   log_dir = Path("logs/tensorboard-demo")
   if log_dir.exists():
       shutil.rmtree(log_dir)
    
   x_train = np.linspace(0.0, 1.0, 80, dtype="float32").reshape(20, 4)
   y_train = ((x_train[:, 0] + x_train[:, 1]) > 0.8).astype("float32")
    
   model = keras.Sequential(
       [
           keras.Input(shape=(4,), name="features"),
           keras.layers.Dense(8, activation="relu"),
           keras.layers.Dense(1, activation="sigmoid"),
       ]
   )
   model.compile(
       optimizer=keras.optimizers.Adam(learning_rate=0.03),
       loss=keras.losses.BinaryCrossentropy(),
       metrics=[keras.metrics.BinaryAccuracy(name="accuracy")],
   )
    
   callback = keras.callbacks.TensorBoard(
       log_dir=log_dir,
       histogram_freq=0,
       write_graph=False,
       update_freq="epoch",
   )
   history = model.fit(x_train, y_train, epochs=2, batch_size=5, callbacks=[callback], verbose=0)
   event_files = sorted(path for path in log_dir.rglob("events.out.tfevents.*") if path.is_file())
    
   print(f"backend: {keras.backend.backend()}")
   print(f"log directory: {log_dir}")
   print(f"epochs completed: {len(history.history['loss'])}")
   print(f"event files: {len(event_files)}")
   print(f"first event file: {event_files[0] if event_files else 'none'}")
   print(f"view command: tensorboard --logdir {log_dir}")

   Use a unique log_dir for every run or experiment variant.

3. Run the script.

   $ python log_tensorboard.py
   backend: tensorflow
   log directory: logs/tensorboard-demo
   epochs completed: 2
   event files: 1
   first event file: logs/tensorboard-demo/train/events.out.tfevents.1783350249.643c51bdba63.453.0.v2
   view command: tensorboard --logdir logs/tensorboard-demo

4. Confirm that an event file was written.

   event files: 1 means the callback created TensorBoard data under the configured log directory.

5. Open TensorBoard against the log directory.

   $ tensorboard --logdir logs/tensorboard-demo

   Use the URL printed by TensorBoard to inspect scalar charts for loss and accuracy.