Keras: use Tensorboard with train_on_batch()

问题:

For the keras functions fit() and fit_generator() there is the possibility of tensorboard visualization by passing a keras.callbacks.TensorBoard object to the functions. For the train_on_batch() function there obviously are no callback available. Are there other options in keras to create a Tensorboard in this case?

回答1:

I think that currently the only option is to use tensorflow code. In this stackoverflow answer I found a way to create a tensorboard log manually. Thus a code sample with the keras train_on_batch() could look like this:

# before training init writer (for tensorboard log) / model
writer = tf.summary.FileWriter(...)
model = ...

# train model
loss = model.train_on_batch(...)
summary = tf.Summary(value=[tf.Summary.Value(tag="loss", 
                                             simple_value=value), ])
writer.add_summary(summary)

Note: For this example in Tensorboard you have to choose the Horizontal Axis "RELATIVE" as no step is passed to the summery.

回答2:

A possible way to create the TensorBoard callback, and drive it manually:

# This example shows how to use keras TensorBoard callback
# with model.train_on_batch

import tensorflow.keras as keras

# Setup the model
model = keras.models.Sequential()
model.add(...) # Add your layers
model.compile(...) # Compile as usual

batch_size=256

# Create the TensorBoard callback,
# which we will drive manually
tensorboard = keras.callbacks.TensorBoard(
  log_dir='/tmp/my_tf_logs',
  histogram_freq=0,
  batch_size=batch_size,
  write_graph=True,
  write_grads=True
)
tensorboard.set_model(model)

# Transform train_on_batch return value
# to dict expected by on_batch_end callback
def named_logs(model, logs):
  result = {}
  for l in zip(model.metrics_names, logs):
    result[l[0]] = l[1]
  return result

# Run training batches, notify tensorboard at the end of each epoch
for batch_id in range(1000):
  x_train,y_train = create_training_data(batch_size)
  logs = model.train_on_batch(x_train, y_train)
  tensorboard.on_epoch_end(batch_id, named_logs(model, logs))

tensorboard.on_train_end(None)