Accessing TensorBoard Data as DataFrames

Overview

The main feature of TensorBoard is its interactive GUI. However, users sometimes want to programmatically read the data logs stored in TensorBoard, for purposes such as performing post-hoc analyses and creating custom visualizations of the log data.

TensorBoard 2.3 supports this use case with tensorboard.data.experimental.ExperimentFromDev(). It allows programmatic access to TensorBoard's scalar logs. This page demonstrates the basic usage of this new API.

Setup

In order to use the programmatic API, make sure you install pandas alongside tensorboard.

We'll use matplotlib and seaborn for custom plots in this guide, but you can choose your preferred tool to analyze and visualize DataFrames.

pip install tensorboard pandas
pip install matplotlib seaborn

from packaging import version

import pandas as pd
from matplotlib import pyplot as plt
import seaborn as sns
from scipy import stats
import tensorboard as tb

major_ver, minor_ver, _ = version.parse(tb.__version__).release
assert major_ver >= 2 and minor_ver >= 3, \
    "This notebook requires TensorBoard 2.3 or later."
print("TensorBoard version: ", tb.__version__)

TensorBoard version:  2.3.0a20200626

Loading TensorBoard scalars as a pandas.DataFrame

Once a TensorBoard logdir has been uploaded to TensorBoard.dev, it becomes what we refer to as an experiment. Each experiment has a unique ID, which can be found in the TensorBoard.dev URL of the experiment. For our demonstration below, we will use a TensorBoard.dev experiment at: https://tensorboard.dev/experiment/c1KCv3X3QvGwaXfgX1c4tg

experiment_id = "c1KCv3X3QvGwaXfgX1c4tg"
experiment = tb.data.experimental.ExperimentFromDev(experiment_id)
df = experiment.get_scalars()
df

df is a pandas.DataFrame that contains all scalar logs of the experiment.

The columns of the