How to put scalars in the same row in tensorboard in TF2.0?

[zhangjh915]

Please make sure that this is a feature request. As per our GitHub Policy, we only address code/doc bugs, performance issues, feature requests and build/installation issues on GitHub. tag:feature_template

System information

• TensorFlow version (you are using): 2.0
• Are you willing to contribute it (Yes/No): Yes

Describe the feature and the current behavior/state.

I did not see related arguments in the documentation. If there is such function, please let me know. :P

The following code produces a tensorboard with 3 different rows. How to put the first two in one row and the second in the other?

with summary_writer.as_default():
    tf.summary.scalar("training_loss", loss, step=batch_index)
    tf.summary.scalar("learning_rate", optimizer.learning_rate, step=batch_index)
with summary_writer.as_default():
    tf.summary.scalar("validation_loss", val_loss, step=batch_index)

Will this change the current api? How?
Probably
Who will benefit with this feature?

Any Other info.

[rmothukuru]

@zhangjh915,
Can you please refer the issue, 7089 and this Stack Overflow Issue and let us know if it helps. Thanks!

[zhangjh915]

@rmothukuru ,
Thanks for the quick reply. I read the solutions in the links you attached but they both seem to not be TF2.0. To my knowledge, there is neither "session.run" or "add_scalars" in TF2.0. To make my question clear, The the code I attached in the question produced the following tensorboard. But I want to put both learning rate and training loss into the same row (not to put learning rate and training loss in the same graph).

[rmothukuru]

@zhangjh915,
Please refer this Tensorboard Link to migrate 1.x code to 2.0. Thanks!

[zhangjh915]

@rmothukuru ,
Sorry but it does not help with my question. The solutions you attached is to plot different curves in the same graph, but what I want is to put different graphs in the same row of category. In the picture I attached above, I want to put the first two plots in the same row and set the row name to something like "training".

[nfelt]

@zhangjh915 the categories are based on common prefixes in the tag names, delimited by slashes like a filename. So e.g. if you emit the summaries with the names tf.summary.scalar("training/training_loss", ...) and tf.summary.scalar("training/validation_loss", ...) it should do what you want.

You can also make this happen automatically by putting them inside a tf.name_scope() block like this:

with tf.name_scope("training"):
  tf.summary.scalar("training_loss", ...)
  tf.summary.scalar("validation_loss", ...)

[david-ngo-sixense]

Hello @nfelt ,

Do you know how to display two scalars on the same graph (for tensorflow 2.0) ?
I want to display train_accuracy/valid_accuracy on the same graph. (and train_loss/valid_loss)
Thank you very much.

[psybuzz]

Thanks for the question. Unfortunately, merging 2 line charts with 1 line each into 1 line chart with 2 lines is not currently supported in the Scalars dashboard. There is a separate, existing feature request for adding this, so I would suggest thumbs-up'ing and following the other thread if you are interested: see #597

[david-ngo-sixense]

Thanks @psybuzz for your answer.
But I don't really understand. If I use the function model.fit or model.fit_generator; I can easily display two line charts (train accuracy and validation accuracy) into the same chart (to check overfitting). But when I implement the training loops, I can't make it.
train_log_dir = logs_dir.joinpath('train/')
train_log_dir.mkdir(exist_ok=True)
valid_log_dir = logs_dir.joinpath('valid/')
valid_log_dir.mkdir(exist_ok=True)
train_summary_writer = tf.summary.create_file_writer(str(train_log_dir))
valid_summary_writer = tf.summary.create_file_writer(str(valid_log_dir))

with train_summary_writer.as_default():
with tf.name_scope("accuracy"):
tf.summary.scalar('train_acc', train_metrics[0].result(), step=epoch)
with valid_summary_writer.as_default():
with tf.name_scope("accuracy"):
tf.summary.scalar('valid_acc', val_metrics[0].result(), step=epoch)

It seems to me that we can do it in tensorflow 1.x

[psybuzz]

I see, in the example code, the train_summary_writer creates a run with a chart named "train_acc", while the valid_summary_writer creates a separate run with a chart named "valid_acc". If the intent is to compare the same "accuracy" metric across these 2 different runs, does changing the "tag name" of the scalar to the same name address the problem?

e.g.

with train_summary_writer.as_default():
  tf.summary.scalar('accuracy', train_metrics[0].result(), step=epoch)
with valid_summary_writer.as_default():
  tf.summary.scalar('accuracy', val_metrics[0].result(), step=epoch)

should produce 1 chart named "accuracy" with 2 lines on it.

[david-ngo-sixense]

Yes, it's working @psybuzz . Thanks very much.