CODEWITHSUNDEEP
tensorflow
Carles Gelada
Carles Gelada

Reputation: 573

How to manually create a tf.Summary()

I often want to log python variables --as opposed to tf tensors.

In the docs it says that "you can pass a tf.Summary protocol buffer that you populate with your own data" but there is no docs for tf.Summary and i could not figure out how to use it.

Anyone knows how to create a Scalar summary this way?

Upvotes: 43

Views: 24876

Answers (3)

mrry
mrry

Reputation: 126154

You can create a tf.Summary object in your Python program and write it to the same tf.summary.FileWriter object that takes your TensorFlow-produced summaries using the SummaryWriter.add_summary() method.

The tf.Summary class is a Python protocol buffer wrapper for the Summary protocol buffer. Each Summary contains a list of tf.Summary.Value protocol buffers, which each have a tag and a either a "simple" (floating-point scalar) value, an image, a histogram, or an audio snippet. For example, you can generate a scalar summary from a Python object as follows:

writer = tf.train.SummaryWriter(...)
value = 37.0
summary = tf.Summary(value=[
    tf.Summary.Value(tag="summary_tag", simple_value=value), 
])
writer.add_summary(summary)

Upvotes: 63

KalenGi
KalenGi

Reputation: 2077

I needed to do many updates to the custom summary variable during training so I implemented mine like so:

Before the loop:

writer = tf.summary.FileWriter(log_folder)
accuracy = None
accuracy_summary = tf.Summary()
accuracy_summary.value.add(tag='accuracy', simple_value=accuracy)

Inside the loop:

if i%20000 == 0:
    accuracy = get_accuracy()
    accuracy_summary.value[0].simple_value = accuracy
    writer.add_summary(accuracy_summary, i)

I'm assuming the indexes to value are in the order in which the variables were added to the summary.

Upvotes: 4

nessuno
nessuno

Reputation: 27042

If you want to log a python value you have to create a placeholder that have to be fed when running the tf.Summary op.

Here's a code snipped

value_ = tf.placeholder(tf.float32, [])
summary_op = tf.scalar_summary("value_log", value_)
my_python_variable = 10
# define everything else you need...
# ...
with tf.Session() as sess:
    for i in range(0, 10):
        sess.run(summary_op, feed_dict={value_: my_python_variable*i})

Upvotes: 6

Related Questions