[Q] Create a similar chart to the gradient one about wandb

[fmamberti-wandb]

Hi @ColdFrenzy - thanks for reaching out!

Are you trying to plot something like this:

If that is the case, you can pass a list containing the distribution over time to wandb.Histogram():

# Generating random data
steps = np.arange(0, 10, 1)
distribution_samples = [np.random.normal(loc=i, scale=1, size=100) for i in steps]

# Log the distribution samples over time using wandb.Histogram
for t, samples in zip(steps, distribution_samples):
    wandb.log({"step": t, "distribution_samples": wandb.Histogram(samples)})

Let me know if you have any questions about this.

[ColdFrenzy]

Hi @fmamberti-wandb thank you for your answer. Actually, if I try to use this method, it actually logs nothing, I'm not sure why.

train_metrics["train_metrics/strategy_duration"] =  wandb.Histogram(np.array([value/sum(strategy_duration.values()) for _, value in strategy_duration.items()]))

That's part of the metric that I log, but doing so, there is no chart called "strategy_duration" on wandb

[fmamberti-wandb]

Hi @ColdFrenzy, how are you logging train_metrics["train_metrics/strategy_duration"] ?

for the Histogram to be plotted as shown above, every time a step is logged (so every time you run wandb.log()) you should pass the array with the distribution for one step to wandb.Histogram(), rather than a list of arrays with all the distributions.

[fmamberti-wandb]

Hi @ColdFrenzy , I wanted to follow up on this request. Please let us know if we can be of further assistance or if your issue has been resolved.

[fmamberti-wandb]

Hi @ColdFrenzy, since we have not heard back from you we are going to close this request. If you would like to re-open the conversation, please let us know!