How to merge not all summaries in tensorflow?

Question

I have big graph with two parts, I run in turn. Both has summaries.

I was calling summaries with node

merged_summary = tf.summary.merge_all()

but noticed, that it causes tensors in second half of graph evalueted before it has sense.

So, how to merge only summaries of one half of my graph?

Answer 1

You can use tf.summary.merge, passing a list of the summaries that you want to merge. For example, if you have the summaries:

cost_summary = tf.summary.scalar('cost_sum', cost)  # for some 'cost' tensor
grad_summary = tf.summary.scalar('grad_sum', grad)  # for some 'grad' tensor

you can merge them by name with:

merged = tf.summary.merge([cost_summary, grad_summary])

So just make merged summary operators for each part of your graph and call them when it makes sense to do so.

Answer 2

assuming you have two lists of summaries of the first and the second graphs, i.e:

summaries_first = [tf.summary.image("my_first_graph_input", image), ...]
summary_second = [tf.summary.scalar("my_second_graph_loss"), ..]

merge each list into a single summary op:

first_graph_summary_op = tf.summary.merge(summaries_first)
second_graph_summary_op = tf.summary.merge(summary_second)

now, whenever you execute a sess.run() on each graph, evaluate it's corresponding summary op and write it.