CODEWITHSUNDEEP
pythontensorflow
user192936
user192936

Reputation:

How do I visualise the loss in a single chart in TensorFlow?

I am trying to train a model with TensorFlow and then obtain a graph of the loss. I am using the below snippet:

for step in range(NUM_BATCHES):
    img, lbl = sess.run([batch_images, batch_labels])
    _, loss_value = sess.run([train_op, cost], feed_dict={X: img, Y: lbl, p_keep_conv: 0.8, p_keep_hidden: 0.5})
    print("Step %d, loss %1.5f" % (step, loss_value))
    sys.stdout.flush()
    tf.summary.scalar('loss', loss_value)
    summary_writer.add_summary(sess.run(tf.summary.merge_all()), step)

When I open the log directory in TensorBoard, I can't see a graph of the loss function against runs. In fact, I don't even see an Events section, but I do have a Scalar section which displays LOSS_xx values (xx for each batch).

What am I missing?

Upvotes: 0

Views: 150

Answers (1)

E_net4
E_net4

Reputation: 29972

The mistake here is that summaries are created with TensorFlow operations, and functions in the tf.summary namespace such as tf.summary.scalar also work with tensors. Your code is creating a new, independent summary chart in each iteration, rather than having all scalar loss values as part of the same summary.

To fix this, you must create the summary operations only once, usually before training takes place. With this approach, you only need one run per step:

tf.summary.scalar('loss', cost)
all_summaries = tf.summary.merge_all()
for step in range(NUM_BATCHES):
    img, lbl = sess.run([batch_images, batch_labels])
    _, loss_value, summary = sess.run([train_op, cost, all_summaries], feed_dict={X: img, Y: lbl, p_keep_conv: 0.8, p_keep_hidden: 0.5})
    print("Step %d, loss %1.5f" % (step, loss_value))
    sys.stdout.flush()
    summary_writer.add_summary(summary, step)

Upvotes: 2

Related Questions