tensorboard() fails with newer tensorboard versions

[dfalbel]

Description

tensorflow::tensorboard() fails with recent versions of TensorBoard because the internal tensorboard_version() function calls system("tensorboard --version_tb", intern = TRUE), which returns character(0) with newer TensorBoard versions (the --version_tb flag no longer exists).

This causes package_version(character(0)) to error with "argument is of length zero".

Traceback

Error in `.f()`:
! argument is of length zero
Backtrace:
    ▆
 1. └─tensorflow::tensorboard(temp, port = 6060)
 2.   └─tensorflow:::launch_tensorboard(...)

Root cause

In launch_tensorboard(), the function calls tensorboard_version() to decide how to format the --logdir argument. tensorboard_version() is:

function () {
    if (is.null(ver <- .globals$tensorboard_version)) {
        ver <- package_version(system("tensorboard --version_tb", 
            intern = TRUE, ignore.stderr = TRUE))
        .globals$tensorboard_version <- ver
    }
    ver
}

The --version_tb flag is no longer recognized by newer TensorBoard, so system() returns character(0), and package_version(character(0)) errors.

Suggested fix

Use tensorboard --version instead of --version_tb, or handle the case where the command returns an empty result.

Environment

• TensorBoard installed via pip install tensorboard (version bundled with TensorFlow 2.18+)
• Observed on Ubuntu (GitHub Actions) and macOS