Tensorflow 2.0 custom_scalar plugin for tensorboard

[PgLoLo]

To set layout for custom_layout plugin in tensorflow 1.x you should use method add_summary in FileWriter class. So the code should look like this:

with tf.summary.FileWriter(LOGDIR) as writer:
   layout_summary = summary_lib.custom_scalar_pb(...)
   writer.add_summary(layout_summary)

In tensorflow 2 add_summary method was removed. How can I set custom_scalar layout now? Is it a bug or a desired API change?

[stephenjfox]

You should be able to follow the instruction in the new tf.summary documentation.

Notedly, the following lines seem to allow you to record whatever metrics you want

# define a metric to track
train_loss = tf.keras.metrics.Mean('train_loss', dtype=tf.float32)
test_loss = tf.keras.metrics.Mean('test_loss', dtype=tf.float32)

# track those metrics
for (x_train, y_train) in train_dataset:
    train_step(model, optimizer, x_train, y_train)
  with train_summary_writer.as_default():
    tf.summary.scalar('loss', train_loss.result(), step=epoch) # this line
    tf.summary.scalar('test_loss', test_loss.result(), step=epoch) # and this line

If the 'loss' and 'test_loss' strings were 'cats' and 'dogs', the API call would still work (it just wouldn't be as helpful to some).
Consider the documentation for writing your own Metric objects if that's your real need.

Hope this helps.

[dcbr]

I managed to make the custom_scalar plugin work in TF2.3 as follows:

import tensorflow as tf
from tensorboard.plugins.custom_scalar import layout_pb2
from tensorboard.plugins.custom_scalar.summary import pb as custom_scalar_pb

writer = tf.summary.create_file_writer(LOGDIR)
layout = custom_scalar_pb(layout_pb2.Layout(...))
with writer.as_default():
    tensor = tf.constant(tf.make_ndarray(layout.value[0].tensor))
    tf.summary.write(layout.value[0].tag, tensor, step=0, metadata=layout.value[0].metadata)

It would however be nice to have native TF2 support for this, allowing us to write something along the lines tf.summary.custom_scalar(layout).

[tilakrayal]

@PgLoLo ,
Is this still an issue?
Could you please update TensorFlow to the latest stable version v.2.6 and let us know if you are facing the same error. Thanks!

[google-ml-butler]

This issue has been automatically marked as stale because it has no recent activity. It will be closed if no further activity occurs. Thank you.

[nfelt]

This is still an open issue; I've transferred it to the TensorBoard repository since that's a better place to track it, since if we were to add a TF 2.0 native summary op for custom scalars, it would be defined in TensorBoard's codebase.

Right now, we haven't worked on the custom scalars dashboard in quite a while due to low usage and a general desire to approach this kind of feature differently (allowing UI-driven configuration rather than requiring the layout to be hardcoded in the written summaries).

At this point, I'd recommend using the workaround from #5343 (comment). (There is also the tf.summary.experimental.write_raw_pb() approach used in the updated demo code as of #5011, if you have a serialized version of the v1 summary proto.)