tf.summary API — Custom Logging

Beyond what model.fit logs automatically

The Keras TensorBoard callback handles standard metrics. For custom metrics, sample images, attention maps, embedding projections, or anything not covered by model.fit, use the tf.summary API directly.

Pattern: create a writer, open it as a context, call tf.summary.scalar / image / histogram / text with a name, value, and step. Use separate writers for train and validation logs so they appear as separate series in the dashboard.

Code

Custom logging — scalar / image / histogram·python

import tensorflow as tf
import datetime

log_dir = "logs/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
train_writer = tf.summary.create_file_writer(log_dir + "/train")
val_writer   = tf.summary.create_file_writer(log_dir + "/val")

# Scalars (run inside the training loop, per epoch)
with train_writer.as_default():
    tf.summary.scalar("loss",           train_loss,    step=epoch)
    tf.summary.scalar("accuracy",       train_accuracy, step=epoch)
    tf.summary.scalar("learning_rate",  lr_schedule(epoch), step=epoch)

with val_writer.as_default():
    tf.summary.scalar("loss",     val_loss,     step=epoch)
    tf.summary.scalar("accuracy", val_accuracy, step=epoch)

# Images (shape: batch, h, w, channels; values [0,1] or [0,255])
with train_writer.as_default():
    tf.summary.image("training_samples", images, step=epoch, max_outputs=5)

# Histograms — watch weights evolve
with train_writer.as_default():
    for layer in model.layers:
        for weight in layer.weights:
            tf.summary.histogram(weight.name, weight, step=epoch)

# Text — log sample predictions, hparams, anything string-shaped
with train_writer.as_default():
    tf.summary.text("sample", "Predicted: cat, Actual: cat", step=epoch)

train_writer.flush()
val_writer.flush()

Custom training loop with summaries·python

import tensorflow as tf

train_loss_metric = tf.keras.metrics.Mean(name='train_loss')
train_acc_metric  = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy')
train_writer = tf.summary.create_file_writer("logs/custom_training")

@tf.function
def train_step(x, y):
    with tf.GradientTape() as tape:
        predictions = model(x, training=True)
        loss = loss_fn(y, predictions)
    grads = tape.gradient(loss, model.trainable_weights)
    optimizer.apply_gradients(zip(grads, model.trainable_weights))
    train_loss_metric.update_state(loss)
    train_acc_metric.update_state(y, predictions)

for epoch in range(NUM_EPOCHS):
    train_loss_metric.reset_state()
    train_acc_metric.reset_state()

    for x_batch, y_batch in train_dataset:
        train_step(x_batch, y_batch)

    with train_writer.as_default():
        tf.summary.scalar("loss",     train_loss_metric.result(), step=epoch)
        tf.summary.scalar("accuracy", train_acc_metric.result(),  step=epoch)

tf.summary API — Custom Logging

Beyond what model.fit logs automatically

Code

Progress

Comments 0