Add zooming to the histogram dashboard

[vitchyr]

It'd be nice if I could zoom on the x-axis for the histogram. One or two outliers will make the entire histogram "zoomed out" and look like this:

I tried changing the bin_edges in the HistogramProto so that each bin has approximately the same count, but that didn't change anything when visualizing the histogram.

[wchargin]

Unfortunately, zooming in won't deal with the problem of outliers. The actual summary data that we store has 30 buckets, whose extent spans the extent of your data—all of it. Outliers in your data will cause the interesting data to be clumped together in one bucket. This happens at TensorFlow summary-writing time; there's nothing that the frontend can do to recover that data. You'll need to preprocess your data to remove outliers before passing it to tf.summary.histogram. (You may want to do this at different η-thresholds, etc.)

[vitchyr]

I thought that setting the number of bins would fix that. Is there any way to modify the 30 bucket limit? I tried looking at the code for HistogramProto, but it's a bit obscure (HistogramProto = _reflection.GeneratedProtocolMessageType('HistogramProto', (_message.Message,), ....).

Are the bin edges completely ignored? I thought this would fix that issue too.

[nikonikolov]

A follow up on the above. The distributions tab alleviates this problem to a degree, but it also does not allow zooming. So for example in the plot below, because the data shrinks very abruptly, it is very hard to track the standard deviation when the data shrinks. It would be nice to be able to zoom in a specific area by drawing a rectangle over it (the same way as in the scalars tab).

[jart]

TensorBoard reservoir samples data to fit in memory. I just submitted a change today that'll allow our SQLite storage system (requires a real disk) to store unlimited samples. After that we can add zooming.

[nikonikolov]

Great news, looking forward to it!

[dmoham1476]

Can we also export histogram values to csv? I see this option is available for scalars

[nfelt]

@dmoham1476 would you mind opening a separate feature request for this, since it's not related to zooming?

[sharvil]

Is there an update on this bug?

[alexbeloi]

@sharvil A temporary workaround until everything is integrated is to clip the values to a known range

instead of

tf.summary.histogram('some_name', my_tensor)

use

tf.summary.histogram('some_name, tf.clip_by_value(my_tensor, min_val, max_val))

where min_val and max_val are your expected lower and upper bounds for values in my_tensor. After this, if you see spikes at the edges of your histogram (as below on the left), that means you have lots of values that are either at that value or being clipped.