I have noticed for year that tensorboard v2.x subscamples scalars when plotting. I never tried to reproduce the problem but noticed if I had a longish training run, every epoch’s loss was not displayed on tensorboard. This would produce a strange jaggy effect of the plot which I never liked, especially given my signal processing background. It was also not obvious why or how it was subsampling. For all I knew, it wasn’t even subsampling but doing some kind of filtering which resulted in the skipping of scalar values on the plots.
Today, I finally figured out how to fix this issue. I was able to correct the issue by using a tensorboard command prompt of:
start “tensorboard” “e:\python_3.8\scripts\tensorboard” –samples_per_plugin scalars=9999999 –logdir .
The addition of the –samples_per_plugin scalars=999999 fixes the issue. Now, all of the points I write out to tensorboard are displayed.
Ref: https://stackoverflow.com/questions/43763858/change-images-slider-step-in-tensorboard