Allow use of histogram visualization with non-histogram summary data

[aidangomez]

I have a use case that I think is pretty common: I want to visualize the categorical probabilities output by a softmax in the form of a histogram.

I.e. I have a vector [0.0 0.5 0.3 0.2] and I want each of those values to specify the height of a bin in a histogram.

Currently there is no way to do this, but I think this is perhaps the most standard modus operandi when preparing data for plotting in a histogram.

[nfelt]

See e.g. #1866 which suggests adding tf.summary.timeseries (though I'm not sure that's the only type of data we could use the histogram visualization for, either).

[ruler501]

Going off the suggestion for timeseries I found a way to implement it using the current APIs. You can probable get rid of the with line if you want to avoid experimental APIs.

import tensorflow as tf
from tensorboard.plugins.histogram.metadata import create_summary_metadata

def log_timeseries(name, values_array, step=None, description=None, display_name=None):
    with tf.summary.experimental.summary_scope(name, "TimeSeriesSummaryOp", [values_array, step]) as (tag, _):
        return tf.summary.write(
            tag=tag,
            tensor = tf.stack(
                [
                    tf.range(0, values_array.shape[0], dtype=tf.float32),
                    tf.range(1, values_array.shape[0] + 1, dtype=tf.float32),
                    values_array
                ],
                axis=1
            ),
            step=step,
            metadata=create_summary_metadata(display_name, description)
        )

EDIT: Okay this doesn't quite work since tensorboard resamples the data it loads into buckets again which requires finding a way around that which looking at the Tensorboard source does not seem possible at this time.

[lminer]

Just echoing that this would be a nice feature

[lukysummer]

Any updates for this plugin?

[nfelt]

No updates to share at this point, I'm afraid.