As we prepare our datasets we are using WandB to track results of different runs. These data-prep runs last several days across 10M+ frames and we would like to log the ongoing results to histograms. For example, we want to generate histograms of statistics metrics across all data frames.
The simplest technique would seem to be to log each parameter step-by-step for each frame using wandb.log(). But the workspace Panel (“Add Panel”) does not not support histograms.
I’ve read the section on histograms repeatedly and I can’t help but conclude that continuous tracking of histograms is not supported. Which seems odd.
The documentation suggests that I should accumulate the elements in a local list, generate a histogram with each frame (locally) and to then log the histogram to WandB. Basically we are just plotting the histogram locally and pushing a figure. This seems inefficient for large data sets.
WandB would seem to be well suited for tracking statistics across steps and runs using histograms.
But perhaps I have not yet found it in the documentation. Can you provide some guidance?
To reiterate, I want to log a several statistics every step across many (millions) of steps. I wish to plot a histogram of these values across all steps.