3

I am using Tensorboard 2 to visualize my training data and I am able to save scalar plots to disk. However, I am unable to find a way to do this for histogram plots (tf.summary.histogram).

Is it possible to save histogram plots from Tensorboard 2 to disk, just like it is possible to do with scalars? I have looked through the documentation and it seems like this is not supported, but I wanted to confirm with the community before giving up. Any help or suggestions would be greatly appreciated.

Daniel Wiczew
  • 165
  • 9

1 Answers1

1

There is an open issue to add a download button for histograms. However, this issue is open for more than 4 years, so I doubt it is getting resolved soon.

A workaround is to use the url that tensorboard would use to get the data. A short example:

# writing some data to tensorboard
from torch.utils.tensorboard import SummaryWriter
import numpy as np

writer = SummaryWriter('./tmp')

writer.add_histogram('hist', np.arange(10), 0)

Open tensorboard in the browser (here localhost:6006):

Tensorboard

Get data as JSON using the template
http://<tb-host>/data/plugin/histograms/histograms?run=<run-name>&tag=<tag-name>.
Here http://localhost:6006/data/plugin/histograms/histograms/?run=.&tag=hist:

Data

Now you can download the data as JSON.
Quick comparison with matplotlib:

import pandas as pd
import json
import matplotlib.pyplot as plt

with open('histograms.json', 'r') as f:
    d = pd.DataFrame(json.load(f)[0][2])

fix, axes = plt.subplots(1, 2, figsize=(10, 3))
axes[0].bar(d[1], d[2])
axes[0].set_title('tb')

axes[1].hist(data)
axes[1].set_title('original')

Hist

Plagon
  • 2,689
  • 1
  • 11
  • 23
  • The problem is, that this way you don't get access to all data (I don't know why). For example there's 200 epochs, but only 50 of them are saved in the data – Daniel Wiczew Jan 27 '23 at 13:51