Suboptimal subsampling behavior

I am trying to build a plot of “loss/eval” vs “completed_steps”. “loss/eval” is the validation loss and it is logged once in a while. “completed_steps” is logged at every step. I would like to see all datapoints where “loss/eval” is logged to be displayed in the plot, because there is not many of them (about 30 for each run). Instead I only see random ones because apparently the downsampling procedure is based on frequency of “completed_steps” logging. As a result the plots are not very informative;


I think a better behavior when plotting a metric X against metric Y would be to fetch all (X, Y) pairs first and then downsample if there is too many such pairs.

Can I hope to see improvements in downsampling logic at some point?

This discussion was followed up via email, but I’ll follow up here once it’s resolved