tfevents

tfevents allows logging data from machine learning experiments to a file format that can be later consumed by TensorBoard in order to generate visualizations.

Installation

You can install tfevents from CRAN with:

install.packages("tfevents")

You can install the development version of tfevents from GitHub with:

You need to have cmake on your path. See installation instructions in the cmake install webpage - or:

If you use brew on MacOS you can run:

brew install cmake

Or on linux install the cmake library, for example on Debian systems:

sudo apt install cmake

# install.packages("devtools")
devtools::install_github("mlverse/tfevents")

Example

The main entrypoint in tfevents API is the log_event function. It can be used to log summaries like scalars, images, audio (Coming soon), histograms (Coming soon) and arbitrary tensors (soon) to a log directory, which we like to call logdir. You can later point TensorBoard to this logdir to visualize the results.

library(tfevents)

Summaries are always associated to a step in the TensorBoard API, and log_event automatically increases the step everytime it’s called, unless you provide the step argument.

Let’s start by logging some metrics:

epochs <- 10
for (i in seq_len(epochs)) {
  # training code would go here
  log_event(
    train = list(loss = runif(1), acc = runif(1)),
    valid = list(loss = runif(1), acc = runif(1))
  )
}

By default this will create a logs directory in your working directory and write metrics to it - you can change the default logdir using context like with_logdir or globally with set_default_logdir().

Since we passed a nested list of metrics, log_event will create subdirectories under logs to write metrics for each group.

fs::dir_tree("logs")
#> logs
#> ├── train
#> │   └── events.out.tfevents.1673447763.v2
#> └── valid
#>     └── events.out.tfevents.1673447763.v2

You can later point TensorBoard to that logdir using TensorBoard’s command line interface or tensorflow’s utility function tensorboard()

tensorflow::tensorboard(normalizePath("logs"), port = 6060)
#> Started TensorBoard at http://127.0.0.1:6060

TensorBoard will display the results in a dashbboard, similar to one you can see in the screenshot below:

You can learn more in the tfevents website.