Seoul
Artificial
Intelligence

Group of professionals from different fields meets twice per month in the heart of Seoul to discuss about AI or implement various AI applications.

meetup.com
seoulai.slack.com
twitter.com
facebook.com
blog.seoulai.com
github.com

Organizers
Martin Kersner
Cinyoung Hur
Adel Sohbi

Announcement
Call for presenters!

Content
Presentations from 2017

Keras and Tensorboard

Martin Kersner, 2018 Feb 06

Keras, as a high-level neural networks API, tries to simplify procedure of training neural networks by introducing numerous constraints. Recently, I have stumbled upon one of them when I tried to visualize training metrics with TensoBoard. Keras offers TensorBoard callback that is able to save and visualize graph, training and test metrics and activation histograms for different layers in model.

keras.callbacks.TensorBoard(log_dir='./logs', histogram_freq=0, batch_size=32, write_graph=True, write_grads=False, write_images=False, embeddings_freq=0, embeddings_layer_names=None, embeddings_metadata=None)

At first, it can seem sufficient, but soon you will need to visualize more or display metric names in more practical format. Experimenting with different optimizers, learning rates and their schedules requires to see actual value of learning rate, however, Keras does not support visualization of learning rate out-of-the-box.

Visualized metric names are same as defined in Keras (e.g. “loss”, “val_loss”, “acc”, “val_acc”), but it would be nice to use the power of Tensoboard and use names in hierarchical manner (e.g. (“train/loss”, “val/loss”, “train/acc”, “val/acc”). TensorBoard can then group metrics to logical cells and generated visualizations become easier to read.

To solve both issues we will use tf.summary module and tf.Session class directly.

Add learning rate to Tensorboard

First, we need to define placeholder that will receive learning rate value and then pass it to tf.summary.scalar. For now, we do not want to define more metrics so we just use apply Tensorflow operation tf.summary.merge_all to collect all defined summaries into one operation.

class TensorboardKeras(object):
  def __init__(self, model, log_dir):
      self.model = model
      self.log_dir = log_dir
      self.session = K.get_session()

      self.lr_ph = tf.placeholder(shape=(), dtype=tf.float32)
      tf.summary.scalar('lr', self.lr_ph)
      self.merged = tf.summary.merge_all()
      self.writer = tf.summary.FileWriter(self.log_dir)

The next step is to get current value of learning rate. With Keras we can access it through Keras backend.

def _get_lr(self):
    return K.eval(self.model.optimizer.lr)

Obtaining learning rate, as shown in code above, has to be done in regular steps. Lambda callbacks allows to execute arbitrary function at certain stages of training:

For our purpose, we will utilize on_epoch_end which expects two positional arguments: epoch and logs. LambdaCallback is later fed to fit or fit_generator method as optional argument called callbacks. Notice that callbacks argument expects list as input. We could define several callback functions that will be exectuted at appropriate times.

callbacks = [LambdaCallback(on_epoch_end=lambda epoch, logs: self.tensorboard.on_epoch_end_cb(epoch, logs))]

In on_epoch_end_cb function, we will evaluate merged summaries operation and feed lr_ph placeholder with current learning rate value.

def on_epoch_end(self, epoch, logs):
  summary = self.session.run(self.merged, feed_dict={ self.lr_ph: self._get_lr() })
  self.writer.add_summary(summary, epoch)
  self.writer.flush()

Rename training and validation metrics in Tensorboard

To update remaining metrics, we have to add another metric placeholders and feed current metric values at the end of every epoch.

Below you can find full example of custom Tensorboard class that is supported by Keras.

class TensorboardKeras(object):
  def __init__(self, model, log_dir):
    self.model = model
    self.log_dir = log_dir
    self.session = K.get_session()

    self.lr_ph = tf.placeholder(shape=(), dtype=tf.float32)
    tf.summary.scalar('lr', self.lr_ph)

    self.val_loss_ph = tf.placeholder(shape=(), dtype=tf.float32)
    tf.summary.scalar('val/loss', self.val_loss_ph)

    self.val_acc_ph = tf.placeholder(shape=(), dtype=tf.float32)
    tf.summary.scalar('val/acc', self.val_acc_ph)

    self.train_loss_ph = tf.placeholder(shape=(), dtype=tf.float32)
    tf.summary.scalar('train/loss', self.train_loss_ph)

    self.train_acc_ph = tf.placeholder(shape=(), dtype=tf.float32)
    tf.summary.scalar('train/acc', self.train_acc_ph)

    self.merged = tf.summary.merge_all()
    self.writer = tf.summary.FileWriter(self.log_dir)

  def _get_lr(self):
    return K.eval(self.model.optimizer.lr)

  def on_epoch_end(self, epoch, logs):
    summary = self.session.run(self.merged,
                               feed_dict={self.lr_ph: self._get_lr(),
                                          self.val_loss_ph: logs["val_loss"],
                                          self.train_loss_ph: logs["loss"],
                                          self.val_acc_ph: logs["val_acc"],
                                          self.train_acc_ph: logs["acc"]
                                          })
    self.writer.add_summary(summary, epoch)
    self.writer.flush()

  def on_epoch_end_cb(self):
    return LambdaCallback(on_epoch_end=lambda epoch, logs:
                                       self.on_epoch_end(epoch, logs))