0

When running multiple logs at the same time, I can see the different training curves side-by-side in different graphs.

However I would like to:

  • See all of the same kind of graph (e.g. loss) together on a single chart

  • See the average and variance/spread of the runs on a single chart.

Thanks!

machinaut
  • 495
  • 2
  • 4
  • 17
  • This may not be a direct answer, but one approach is to extract data from the event file using `tf.train.summary_iterator`, then make your own plot with that data. See this thread for more info: https://stackoverflow.com/questions/37304461/tensorflow-importing-data-from-a-tensorboard-tfevent-file/37411543. – Richard_wth May 23 '18 at 12:01

2 Answers2

2

Aggregate charts

If your runs are using the same summary tags (e.g. tf.summary.scalar('loss', loss)) and are saving their events file in subdirs of the same root dir (e.g. /model_results/run_1/, /model_results/run_2/, ...), then by opening Tensorboard on the root dir (tensorboard --logdir=/model_results/) you should be able to see your plots aggregated by tags (at least for recent versions of Tensorboard).

Display average/variance over runs

This seems like a duplicate of How to display the average of multiple runs on tensorboard. @Alex's solution there is quite thorough, using EventAccumulator to combine the scalar summaries over several files and tf.Summary() to write back the accumulated values (mean, variance, etc.).

benjaminplanche
  • 14,689
  • 5
  • 57
  • 69
0

I had the same issue while tuning a Keras model with tensorflow as backend. You can check the issue and the code here: How to use Keras TensorBoard callback for grid search

paolof89
  • 1,319
  • 5
  • 17
  • 31