When I use the histogram function from wandb and collect histograms over time I get a superposition of blurry histograms, which is un-usable (I’m not really sure what I am looking at).
What I would expect to see instead for a multi-histogram visualisation would be like the plot from tensorboard (see above), the offset accross timesteps is important. What’s even more bizarre is that it seems that these plots are already plotted when I hover over the graph in wandb (where I see a clear histogram with a fitted curve above for every timestep I collected).
Would it be possible to visualize those with an offset like in tensorboard? I really want to visualize the evolution of distribution of a value across episodes.
Hi @yann-berthelot thanks a lot for writing in about this. This might be feasible with one of the following ways:
If you are using torch.nn modules, just pass them to wandb.watch
If you want to log tensorboard histograms from your code, you could use wandb.init(sync_tensorboard)=True and wandb wil convert the tensorboard histograms into wandb.Histogram objects
You could pass a tensor of gradients or parameters to wandb.Histogram and wandb.log it in your code.
Please let me know if you have tried any of these, and if that would work for you? Also, it would help to understand what type of data (and training framework) you would like to visualise in this way, so that I could provide a more customized solution if the above aren’t applicable in your case.
Hi @yann-berthelot I wanted to check in here and see if any of the above suggestions would work for you, or if that won’t give you the same functionality? Please also let me know if you have any additional questions.
Hi @yann-berthelot since we haven’t heard back from you in a while, I will go ahead and close this ticket for now. However please let us know here if the above wouldn’t work for you or if you had further questions about it, and we will be happy to assist you further!
